Vision gets obscured in adverse weather conditions, such as heavy downpours, dense fog, haze, snowfall, etc., which increase the number of road accidents yearly. Modern methodologies are being developed at various academics and laboratories to enhance visibility in such adverse weather with the help of technologies. We review different dehazing techniques in many applications, such as outdoor surveillance, underwater navigation, intelligent transportation systems, object detection, etc. Dehazing is achieved in four primary steps: the capture of hazy images, estimation of atmospheric light with transmission map, image enhancement, and restoration. These four dehazing procedures allow for a step-by-step method for resolving the complicated haze removal issue. Furthermore, it also explores the limitations of existing deep learning-based methods with the available datasets and the challenges of the algorithms for enhancing visibility in adverse weather. Reviewed techniques reveal gaps in the case of remote sensing, satellite, and telescopic imaging. In the experimental analysis of various image dehazing approaches, one can learn the effectiveness of each phase in the image dehazing process and create more effective dehazing techniques.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

PP and VD either did not respond directly or could not be reached.

Cloud computing (CC) technologies become popular due to their characteristics of maximum availability, flexibility, and scalability. Despite the benefits of CC, security becomes a major challenging issue that can be resolved by the use of image encryption techniques. Numerous works have been carried out with several approaches, and effective models need to be designed to enhance secure image encryption algorithms. Recently, chaotic approaches have been available in various domains, such as design of cryptosystems and image encryption. Chaos-based methods utilize random chaos sequences for image encryption, which are highly secure and fast. In this view, this study introduces a unique mayfly optimization with chaos-based image encryption (MFO-CBIE) technique for CC environment. The major intention of the MFO-CBIE model is to accomplish security in the CC environment via encryption and key generation process. The proposed MFO-CBIE model follows a three-stage process, namely image encryption, optimal key generation, and image decryption. In addition, the MFO-CBIE model employs the MFO algorithm for the optimal key generation process of the chaos-based encryption technique. Once the images are encrypted, they are securely transmitted to the cloud server. On the receiver side, the cipher image gets accessed from the cloud server, and then the decryption process is carried out. The experimental evaluation of the MFO-CBIE model is tested using benchmark images, and the results ensured the promising performance of the MFO-CBIE model over recent approaches.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of compromised peer review and determined the paper is out of the scope of the journal. The Editor and publisher no longer have confidence in the results and conclusions of the article.

MS disagrees with the retraction. SD either did not respond or could not be reached.

The Editor-in-Chief and the publisher have retracted this article, which was submitted as part of a guest-edited special section. An investigation uncovered evidence of systematic manipulation of the publication process, including compromised peer review. The Editor and publisher no longer have confidence in the results and conclusions of the article.

AT and SR did not agree with the retraction.

Guest editors Yang Li and Sezai Ercisli introduce the Special Section on Advanced Image Processing and Applications for Smart Agricultural Systems.

Object detection and target localization are important technologies in image processing and computer vision, which have a wide range of applications in agricultural systems. By fusing the YOLOv5 algorithm and monocular vision-based method, a target localization algorithm is proposed to accurately identify and locate the various objects in agricultural scenes. The GPS information of the target object can be obtained by means of processing the attitude angle information of the unmanned aerial vehicle (UAV) at the time when it takes pictures. The superiority of the proposed method is demonstrated through the test on the agricultural scenario dataset taken by UAV, with the experimental results showing the algorithm’s satisfactory speed and accuracy.

Farmers are often annoyed by crop pests, which can lead to serious losses if not detected in time. Deep learning has gradually penetrated into the agricultural field due to its proven superior performance, promoting the development of smart agriculture. Its real-time performance and effectiveness can help farmers spot pests in time and reduce losses. However, with the increasing application of deep learning technology, its disadvantage of relying heavily on large amounts of data has gradually been exposed. Before training a network, people need to spend a lot of manpower, time, and money to prepare a complete dataset and make the correct labels. To solve this problem, we introduced a less-shot object detection method in crop pest detection. We provide the first few-shot pest detection dataset (FSPD), which contains 20 categories and 2000 samples for this research. We use a few-shot object detection network based on an object pyramid called multiscale positive sample refinement. Numerous experiments are conducted to prove that, with only few (1/2/3/5/10-shot) samples, the accuracy of our method on FSPD is better than that of traditional deep learning methods, such as YOLO and faster R-CNN. After embedding the object pyramid branch, the network performance improves by an average of 5.9% on three splits compared with that without embedding. With input of only 10-shot novel samples, our method can achieve 76%, 79%, and 82% for the mAP on three splits, respectively. As a research study on the detection of crop diseases and pests with less radiation, our work can give some new insight to the future scholars, which is of great significance to the development of smart agriculture.

Weeds have seriously impacted crop planting and caused serious human losses. The intelligent weed identification algorithm based on the convolutional neural network has made some achievements in intelligent agriculture. However, the current results rely on a large number of labeled weed data, which is costly to obtain. How to effectively identify without big data is a key task in the current field. We designed a high-information data-centric weed images data identification system based on a triple filter. The system consists of three modules: nearest-neighbor core metric, unobserved components model, and outlier detection. Extensive scientific experiments have shown that our system requires only a small amount of data to achieve excellent performance in weed identification tasks. Our system saves 5% to 20% of the data. When using only 80% of the data, our system achieves the model performance obtained with 100% of data training. Compared with other methods, this method improves the accuracy by 4.9% and reaches a state-of-the-art performance. Our work solves the problem of relying on image data in smart agriculture, provides a scheme for weed identification tasks, and provides valuable ideas for future intelligent agricultural research.

Unmanned aerial vehicles (UAVs) are one of the most promising technologies for weed detection and management because they are not constrained by ground conditions, can be easily maneuvered, and can cover a large agricultural area. For effective use in a UAV-based system, the weed detection system must meet stringent requirements regarding power consumption, overall detection performance, real-time capability, weight, and size. We evaluate the three lightweight deep learning architectures, YOLOv4-tiny, YOLOv4-tiny-3l, and YOLOv7-tiny, for real-time weed detection in horticulture, specifically forest tree nurseries. The deep learning models are evaluated on the three edge computing devices, NVIDIA Jetson Nano, NVIDIA Jetson TX2, and NVIDIA Jetson Xavier NX, that fit the requirements for UAV-based applications. The deep learning models are trained and evaluated with a custom dataset of 921 images and 52,081 individual weed and horticultural plants under normal commercial growing conditions. The deep learning architecture YOLOv4-tiny-3l achieves the highest F₁-score with 80.56% and has the highest average precision for weeds and three of the five horticultures. Moreover, an mAP@0.5 of 83.38% and mAP@0.5:0.95 of 46.89% were attained. The NVIDIA Jetson Xavier NX achieves the highest detection speeds for all image input sizes, with over 60 FPS at FP16 and over 100 FPS at INT8 for an image input size of 832 × 832 pixels, making it suitable for use in UAV-based real-time weed detection in horticulture.

The frequent outbreak of crop pests is one of the main factors affecting crop yield. The accurate identification of pests is significant for effective pest control, which is the basis for the safe growth of crops and is also an important guarantee for high crop quality and crop yield. Computer vision based on supervised deep learning enables intelligent identification of pests, whose success is still inseparable from a large amount of labeled data, causing a large number of resource consumption due to data labeling. Therefore, an in-depth study on maximizing the value of data is essential when lacking labeled data. A new data evaluation method based on Mahalanobis distance and entropy is proposed to address the problem of lacking labeled data in intelligent pest identification. This method enables filtering high-value data, thus achieving effective pest identification performance with a small data volume. The experiment is conducted on a dataset we collected called PD-20, which shows the proposed method achieves baseline accuracy of 100% using only about 60% of the original data. Moreover, the proposed method can save at least 10% of the data volume compared with three comparison methods. To facilitate the deep integration of smart agriculture with artificial intelligence (AI), we designed an interactive framework of active learning for pest identification based on the proposed method, which lays the foundation for the application of AI in direction of agriculture.

Plants are crucial in providing sufficient food for the increasing global population. To be able to provide an appropriate amount of food, the maximization of agricultural output is needed while input needs to be minimized. For these purposes, plant phenotyping techniques, i.e., measuring and analyzing the physical and biochemical characteristics of plants, can be employed. One of the most essential indicators of the general health and development of the plant is the color, shape, and number of leaves. To analyze plant images and capture essential plant traits, various algorithms have been developed. However, one of the important challenges in developing these algorithms is the occlusion or overlapping of leaves and biomass. We present a solution for leaf counting in the presence of occlusion in the plant Arabidopsis thaliana that includes four different convolutional neural network architectures. Datasets from the Computer Vision Problems in Plant Phenotyping (CVPPP) 2017 challenge and Photon System Instruments were used. The results are discussed in detail and compared with the existing solutions. Results showed that our solutions for leaf counting are superior to the previous winners of the CVPPP challenges.

Reversible data hiding in encrypted images (RDHEI) is a well-established technique in which additional information is embedded to ensure data imperceptibility and full reversibility. For RDHEI, how to utilize more redundant space to improve the embedding capacity (EC) under the premise of true reversibility is still a challenging task. We propose a separable RDHEI scheme based on multiple predictive compression coding. The image is classified by texture complexity based on gray level co-occurrence matrix. For different blocks, adaptive weight prediction (AWP) and global optimization adaptive weight prediction (GOAWP) are proposed to improve prediction accuracy. More importantly, GOAWP solves the problem of inaccurate prediction of texture block. Furthermore, methods based on pixel difference compression, pixel bit-based difference compression, and block plane-based difference compression are proposed to vacate more embedding space. The proposed scheme is separable. For the content owner, above two predictors and three compression methods are adopted to vacate space for hiding. For the data hider, the secret bits are embedded into the encrypted image by directly replacing the spare bits without obtaining any information of the original image. For the receiver, the desired information can be retrieved according to the key in his/her possession. By taking full advantage of the pixel correlation and multiple predictive compression coding, the proposed method successfully achieves a significant improvement in capacity. Experimental results show that the highest payload of our scheme reaches 3.6185 bpp, which exhibits appreciated EC in comparison with state-of-the-art methods. Security and compression performance is emphasized, where the key space of our scheme is as high as 2 × 10¹⁵ × ∞ × 10¹⁵, the average information entropy of the embedded compressed image is 7.8891, and the correlation coefficients between original images and the compressed and embedded compressed images are close to 0, indicating a weak correlation, thereby greatly improving security.

We aim to enhance the ConvNext model using high-order spatial interactions and structural reparameterization techniques. By comparing with current mainstream convolutional neural networks and Transformer models, a more efficient model structure named Rep2former was designed. The proposed model retains efficient spatial interaction capabilities while avoiding the problem of quadratic complexity, which is achieved by not using self-attention. Furthermore, a balanced adjustment was made in parameter quantity, floating-point operations per second, and accuracy to ensure that the model is suitable for deployment in applications. Experimental evaluations and comparisons were conducted on the ImageNet-1k and CIFAR-10 datasets. The results indicate that the Rep2former model outperformed popular models such as ResNet-50, Swin Transformer, and ConvNext in terms of parameter quantity, computational complexity, and accuracy. Therefore, the proposed model is considered more effective and user-friendly.

Generating action videos in future scenes based on static images can make computer vision systems to be better applied for video understanding and intelligent decision-making. However, current models pay more attention to the motion trend of the generated objects, and the processing effect on local details is not ideal. The local features of the generated video will have the problem of blurred frames and incoherent motion. This paper proposes a two-stage model, video spatio-temporal generative adversarial network (VSTGAN), which consists of two GAN networks, such as temporal network and spatial network (S-net). The model fully combines the advantages of CNNs, recurrent neural networks (RNNs), and GANs to decompose the complex spatiotemporal generation problem into temporal and spatial dimensions. Therefore, VSTGAN can focus on local features from the above dimensions respectively. In the temporal dimension, we propose an RNN unit, the convolutional attention unit (ConvAU), which uses the convolutional attention module to dynamically generate weights to update the hidden state. Thus, T-net uses the ConvAU to generate local dynamics. In the spatial dimension, S-net uses CNNs and attention modules to perform resolution reconstruction of the generated local dynamics for video generation. We build two small-sample datasets and validate our approach on these two new datasets and the KTH public dataset. The results show that our approach can effectively generate local details in future action videos and that the model performance on small-sample datasets is competitive with the state-of-the-art in video generation.

Oral cancer is a serious worldwide health problem. It might be seen in the face, oral glands, neck, or mouth of the patient. Cancer detection utilizing histopathology images aids in easing and forecasting abnormality. Furthermore, it delivers better outcomes if biological procedures are implemented appropriately, however, there are large opportunities for human errors and blunders during physical examinations. The advancement of deep learning techniques may open the door to a better detection of oral cancer from histopathology images, which will be advantageous to lab staff and medical professionals. Our study presents an in-depth analysis of 10 pre-trained deep convolutional neural network models using transfer learning approach for the detection of oral cancer. The experiment was performed for two classes, i.e., normal and oral squamous cell carcinoma. The results show that the VGG19 model with data augmentation was able to attain the highest classification accuracy of 96.26% using the transfer learning technique. We have also introduced an approach by integrating the VGG19 pre-trained model with a custom naïve inception block. This fusion of two well-known models harnesses their complementary strengths and results in a more robust architecture. The incorporation of the inception block addresses limitations observed in the VGG19 framework, such as vanishing gradient issues and excessive computational requirements. The proposed model for detection of oral cancer has undergone a thorough block-wise fine tuning. Our results demonstrate the superior performance of our deep learning architecture compared to existing literature, highlighting its potential to enhance the detection and diagnosis of oral cancer.

In recent years, finger vein (FV) recognition has garnered significant attention due to its inherent advantages, such as enhanced security, convenience, and the ability to discern living organisms. Notably, use of vision transformers in FV recognition has yielded promising results, primarily owing to their adeptness in capturing extensive spatial relationships within images. Nevertheless, transformers necessitate augmented computational resources and presently fall short compared to established convolutional neural networks (CNNs) concerning performance metrics. We address this predicament and propose the development of an advanced network that surpasses conventional transformers and convolutional networks. By leveraging transformers to capture long-range dependencies and CNNs to extract localized information, we introduce a hybrid architecture named the FV convolutional transformer for FV identification. We validate the efficacy of our approach by conducting extensive experiments on three publicly available FV databases. The experimental results demonstrate that our network achieves state-of-the-art performance, as evidenced by the attainment of the lowest equal error rate across all three datasets

To further improve the feature representation ability of haze removal model on small-scale datasets, we propose a three-hierarchical encoder–decoder integrating learning architecture. Unlike increasing the depth of network to improve the performance, we use three sets of transfer learning encoders to aggregate features on multiple patches. This strategy makes full use of deep and shallow features from different spatial parts of original hazy images. These different spatial parts are used to jointly optimize non-homogeneous hazy images with different scene depths and degradation degrees. We propose an integrated learning decoder, in which the internal integrated learning module decodes by gradually infusing the deep and shallow features of the identical patch. We fuse all encoded features at each stage for original resolution decoder. We report the quantitative and visual results of eight advanced haze removal methods (2019 to 2021). We compare these methods on five real-world and synthetic hazy datasets, including homogeneous and non-homogeneous degraded datasets. We present the results of the ablation study for key components. The generalization of our method is significantly better than other methods on real hazy scenes. We report improvements in semantic segmentation performance. This result could be used for autonomous driving, automated guided vehicle scheduling in closed factory area, etc.

Most existing multi-object tracking models only focus on the spatial information generated by image-level input while ignoring the necessary temporal information. The temporal motion information between consecutive frames can effectively reflect the target’s motion status, which is essential to improving the performance in dealing with occlusion and motion blur of the model. We propose MISTracker to realize the motion information supplement to the original tracking model. Specifically, we divide multi-scale feature maps into two categories from the perspective of space and channel information. Meanwhile, the spatial-level frame differences processing (SFDP) module and the channel-level frame differences processing (CFDP) module are proposed to deal with these differences between continuous frames, respectively. The SFDP processes the differences from the perspective of spatial information and supplements the motion information through the perception of pixel-level information changes in the feature maps. The CFDP processes the differences from the perspective of channels and enhances the information of motion-sensitive channels through the overall pixel differences of different channels. Eventually, temporal and motion information are complementary to each other after upsampling fusion. The whole process is realized by simple convolution, which reduces the computational force as much as possible and enhances the tracking performance of the model.

Deep neural networks achieve significant progress in human pose estimation (HPE), but there are still many difficulties in practical applications. One of the important reasons is that the current advanced models are very complex leading to huge computation cost. We propose a flow-based attention lightweight network (FALNet) for HPE to solve this problem. We first design a cheap lightweight bottleneck to reduce the model size and complexity with two components: depthwise convolution and cheap unit. Then we propose a flow-based fusion attention block to generate and aggregate multi-scale features of the same layer effectively. We demonstrate the effectiveness of our methods on two benchmark datasets: the COCO dataset and the MPII dataset. Our network FALNet-50 only has 2.2 M parameters and 0.66G floating-point operations (FLOPs), achieving comparable or even better accuracy with smaller complexity. Moreover, we show the speed advantage of our network during inference. Specifically, our FALNet-50 achieves 70.4 in average precision score on COCO val2017 with 16 FPS inference speed on a smartphone.

Except for recovering the image detail texture information, the main difference between video super-resolution (VSR) and single-image super-resolution (SR) is that VSR focuses on alleviating the deficiency of temporal coherence between video frames. Motion estimation and motion compensation is the common technique used to strengthen the temporal correlation between frames. Most motion estimation methods are based on optical flow. The optical flow method has three basic assumptions: the movement scale is small; the luminance channel is constant; and every pixel in the same image has the same moving trend. In some scenes with complex motion, the accuracy of the underlying optical flow estimator is limited, which leads to artifacts in the video reconstruction. In recent years, generative adversarial network (GAN) has been widely used for VSR reconstruction, which can acquire more realistic texture details for single frame reconstruction. Based on the above reasons, we explore a GAN-based VSR method by optical flow-free motion estimation and compensation [optical flow-free generative and adversarial network (COFGAN)], which completes motion estimation by producing temporal dimension. COFGAN develops better motion estimation result and improves the performance of VSR without optical flow. To verify the motion estimation effect in complex scenes, long-term sequence real dynamic scene dataset realistic and dynamic scenes is applied for training and testing. We compare the performance of proposed COFGAN with earlier works such as video enhancement with task-oriented flow (TOFlow), frame-recurrent video super-resolution (FRVSR), learning temporal coherence via self-supervision for GAN-based video generation (TecoGAN), and so on. Our method achieves significant performance in the temporal coherence metrics performance of learning perceptual image patch similarity (tLP) (0.47) and performance of optical flow estimation (tOF) (7.07) with ×4 up-scaling factor. Compared to the best performance method TecoGAN in previous work, the proposed method promotes 29% of tLP and 26% of tOF. Moreover, COFGAN reaches the best accuracy on the commonly used video sequence datasets Vid4 and ToS3.

It is challenging to remove rain-streaks from a single image because the rain-streaks are spatially varying. Although data-driven rain removal methods have reported promising performance recently, there are still some defects, such as data dependency and insufficient interpretation. A single image deraining algorithm that combines model-based and data-driven approaches is introduced. First, high-frequency information extracted by an improved weighted guided image filter (iWGIF) is used to learn the rain-streaks to avoid interference from other information through the input image. Then, the input and the learned rain streaks are transferred adaptively from the image domain to the feature domain to learn useful features for high quality image deraining. Finally, networks with multi-scale attention mechanisms are used to restore high-quality images from the latent features. Experiments show that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both qualitative and quantitative measures.

Seed phenomics is a comprehensive assessment of complex seed traits, and seed classification is an indispensable step. Plant seed recognition is of great significance in agricultural production, ecological environment, and biodiversity. However, some traditional artificial plant seed classification methods are expensive, time consuming, and laborious. Therefore, there is a need that cannot be ignored for a method to improve the situation. Artificial intelligence is making a huge impact on various fields through its perception, reasoning, and learning capabilities. A challenge in pratacultural research, the rapid auto-identification of plant seeds, might be better resolved by the integration of computer vision. For the lack of a public seed dataset for the training of models, we established a dataset called LZUPSD, which includes images of 88 different species of seeds. We explored methods to achieve fine-grained seed classification using convolutional neural networks and tried to apply a transformer to it. The method has the highest accuracy of more than 95%. The method is able to identify plant seeds automatically with high speed, low cost, and high accuracy. It results in a more efficient plant seed recognition method. At the same time, we have established a platform where users can upload pictures to obtain seed information. In addition, our dataset will be released to the public in the next phase in order to share with interested researchers.

In the realm of picture forensics, it might be difficult to find and locate an image-splicing forgery. To improve the accuracy of the picture forensic evaluation, we introduce a dual encoder network (DAE-Net) with an efficient channel attention (ECA) module. The ECA module creates a fusion approach with an attention mechanism that enables the model to concentrate on local objects’ tampering characteristics and increases the accuracy of multi-region tampering identification. We suggest combining a dual-coding network with a multi-scale dilated convolutional feature fusion module to better detect small target tampering zones. Experimental evidence suggests that DAE-Net outperforms state-of-the-art methods. The attack experiments also demonstrate the DEA-Net model’s stability and noise resistance.

Circle detection in images is one of the key technologies in machine vision, pattern recognition, and artificial intelligence. However, conventional circle detection methods are sensitive to complex scenes, noise, and occlusion in images. Solving the impact of these situations is still the focus of circle detection algorithm research. Therefore, a circle detection algorithm based on neighborhood density clustering (NDC) is proposed. The proposed algorithm calculates circle parameters of connected regions after extracting corners, corroding the corners, and marking the connected regions. Then, NDC of the circle parameters is executed to classify arcs belonging to the same circle into one category to acquire a virtual connected region. And the circle parameters of the virtual connected region are clustered using NDC again to obtain the circle parameter dataset of the circle to be detected. The precise circle parameters are further estimated by calculating the centroid of each category. To prevent false positives, candidate circles are verified through a ratio rule. Extensive experiments using both synthetic and real images were performed. The results compared with those of representative state-of-the-art methods demonstrate that the proposed algorithm can be applied to a variety of complex scenes and has several advantages: good anti-occlusion effect, more robustness against noise, high accuracy, and better performance.

Finger vein recognition is widely used in various fields due to its convenience, high safety, and stability of the finger vein. Because of the uniqueness of biometric characteristics, biometric data protection and privacy raise significant concerns. In addition, studies have shown that user information can be stolen by illegal operations and stored in the system. Once this happens, the user’s information will be leaked everlastingly, and the personal finger vein traits will no longer be unique. To solve this problem, a new cancellable encryption template was proposed for finger vein images, which applies double random phase encoding technology to finger vein image partial mutual encryption. In addition, an adaptive finger vein pattern extraction algorithm based on maximum curvature was proposed to ensure recognition performance while maintaining the security of the template of the user. The SDUMLA and FV_USM finger vein datasets were used for testing to verify the effectiveness of the method. The experimental results show that the performance of our proposed encryption satisfies the criteria of irreversibility, revocability, and unlinkability of biometric template protection and can effectively defend against potential attacks.

Light detection and ranging (LiDAR) camera systems are becoming increasingly vital for autonomous driving. The monocular three-dimensional (3D) detection task is a critical and challenging aspect of this field. However, most algorithms rely solely on manually labeled images, which is a time-consuming and labor-intensive process, and the resulting detection lacks depth information. To address this problem, a semi-supervised 3D object detection model based on LiDAR camera systems (PVONet) is proposed to improve both the detection accuracy and processing time. First, an innovative data preparation block point-voxel fusion estimation is introduced; it utilizes LiDAR points to generate 3D bounding boxes for unlabeled data, thereby significantly reducing the time compared with manual labeling. Second, a new block based on fully connected neural network for box estimation (feature extraction and 3D object detection) is presented; it is utilized to conduct feature extraction, feature correlation, and 3D box estimation on monocular images. Finally, comprehensive experiments conducted on the popular KITTI 3D detection dataset demonstrate that our PVONet is faster (30 ms on KITTI benchmark) and more accurate [with increases of 4.69%/3.82% (easy), 4.45%/2.79% (moderate), and 4.07%/3.75% (hard) aggregation processes on 3D/bird’s eye view objects compared with the baseline]. This meets the requirements for high real-time performance in autonomous vehicles applications. The results demonstrate the effectiveness of our model based on LiDAR camera systems.

The research on sea–land segmentation of remote sensing images has received tremendous attention, which is of great significance to coastline extraction and ocean monitoring. In recent years, various convolutional neural networks (CNNs) have been presented to achieve precise and efficient sea–land segmentation effect. However, existing CNNs typically adopt the symmetric encoder-decoder structure, which is inefficient for feature extraction, feature fusion, and information transmission. To address these problems, this work develops a CNN for pixel-level sea–land segmentation, termed HeteroNet. The proposed HeteroNet constructs a heterogeneous encoder–decoder structure consisting of successive dense-connected encoding modules and squeeze-and-excitation-connected decoding modules that can effectively enhance the feature extraction and fusion capabilities of the network. In addition, an easy-to-embed global context enhanced module is designed to further facilitate information transmission efficiency. Comparative experiments with state-of-the-art methods are conducted to reveal that the HeteroNet can exhibit superior sea–land segmentation performance in different scenarios, and the ablation study is performed to demonstrate the effectiveness of each component in the network.

Composition, color, and brushstrokes are the primary factors in evaluating and appreciating artwork and are vital points to be considered and addressed in arbitrary style transfer tasks. However, existing methods do not balance these elements well, so we propose an approach that balances content structure and style patterns for universal style transformation (BcsUST). Specifically, two lightweight encoders based on residual networks pass the extracted style and content features into the multi-domain structure attention module, which applies a self-attention mechanism to align the input features point by point followed by manifold alignment. A secondary alignment strategy is used to integrate global styles harmoniously into the semantic structure of the content, while obtaining content details to keep the structure intact. Then, the texture modulator generates style convolution parameters to dynamically adjust the filter within the micro-decoder feature subspace, injecting complex and flexible style signals for stylization to enrich the colors and form exquisite strokes. In addition, this work proposes a two-stage training strategy that introduces an artistic appreciation loss to further balance content structure and stylistic signals. Numerous qualitative and quantitative studies demonstrate that the BcsUST framework produces images that resemble the artists’ paintings more closely.

Person verification based on face detection and face recognition is a very important research area in the field of computer vision as it provides authentication before permitting access to resources ensuring safety and security. It is a challenging task to identify and verify a person in low-illumination images, this is because the facial features of a person in a low-illumination image are not clear as the image is of poorer quality than that of an image taken with good illumination. The existing hand-crafted feature-based approaches and deep learning models for low-illumination image contrast enhancement are typically unsatisfactory in the applications of person verification either due to over-enhancing of the image or restricting the contrast of the image while dealing with light illumination. To achieve more accurate face detection and face recognition in low light images, a new approach based on adaptive gamma correction and deep learning model is proposed in this research paper. In this work, two methods: feature-based adaptive gamma correction (FAGC) and deep learning-based adaptive gamma correction (DLAGC) are proposed for contrast enhancement. The proposed approach uses the new adaptive gamma correction-based methods (FAGC, DLAGC) for the image contrast enhancement and applies deep learning models to detect and recognize the face in the enhanced image. The enhancement of the brightness difference between objects and their backgrounds achieved by the proposed adaptive gamma correction-based methods enables the deep learning model to extract the quality semantic information, which improves the accuracy of person verification. The proposed approach is evaluated on Extended Yale Face (EYF) dataset, which is a low-illumination image dataset. The proposed framework with FAGC and DLAGC for person verification achieves an improvement of 24% and 30%, respectively, on EYF dataset and 2.5% and 10%, respectively, on Specs on Faces dataset when compared to the existing techniques.

We aim to devise a technique to share k number of secret images to n number of participants with ( n ≥ k ) in such a way that only k or more participants can collaborate to get the original k secret images. The proposed approach relies on the development of a spanning tree from a complete graph G for the share image generation in which each of the k pixels retrieved from the k secret images is considered to be one of the k vertices of G. Each of the spanning trees of G is represented by a unique sequence of numbers known as Prufer sequences, and each participant is assigned a random Prufer sequence. The Prufer sequence of a participant is used to generate the share images for that participant. The share images are kept in a blockchain using interplanetary file system. This scheme is suitable for protecting sensitive images through distributed storage while eliminating the single points of failure. Because the blockchain is an immutable decentralized public record, anyone can verify their shares as well as the shares of other participants, thereby preventing cheating by any participant. The scheme presents a lossless reconstruction of k secret images, making it perfect for use in military applications, medical applications, and other applications in which lossless recovery is a crucial element. Extensive experimental analysis confirms that the proposed scheme is secure, cheat proof, and lossless.

Biometric feature-based personal identification is gaining popularity these days since it is more trustworthy than previous approaches and has a wide range of applications. Furthermore, hand-based biometrics have better user acceptability, making finger knuckles a relatively acceptable trait due to their location and the added benefit of being less susceptible to harm. Finger knuckle prints are the inherent skin patterns that form at the knuckles of the back of the hand and have been proven to be extremely rich in textures, hence they may be utilized to uniquely identify a person. Our study seeks to explore the innate advantage of finger knuckles and the benefit of combining multiple features from knuckle areas, minor and major, using symmetric sum algorithms to fuse scores from each of the traits. The proposed methodology is tested on a modified AlexNet model as well as the original AlexNet, modified ResNet50, binarized statistical image features, and principal component analysis for comparison. Experimental results are obtained using PolyUKnuckleV1 finger knuckle datasets provided by Hong Kong Polytechnic University. The results reinforced the idea that multimodal biometric systems are stronger and that finger knuckles are unique to each person.

In recent years, research on distracted driving behavior recognition has made significant progress, with an increasing number of researchers focusing on deep-learning-based algorithms. Aiming at the problems of the existing distracted driving recognition algorithm, such as its oversized model and difficulty in adapting to low computing environments, a lightweight network MobileNetV2, is chosen as the backbone network and improved to design a distracted driving behavior detection method that is both accurate and practical. The Ghost module is employed to replace point-by-point convolution to reduce the computation, the Leaky ReLU function helps mitigate the problem of dead neurons, as it prevents gradients from becoming zero for negative inputs. Finally, the channel pruning algorithm is used to further reduce the model parameters. The experiment results on the State Farm dataset show that the model’s test accuracy can reach 94.66%, and the number of parameters is only 0.23 M. The improved model has significantly fewer parameters than the baseline model, which demonstrates the effectiveness and applicability of the method.

The object detection algorithm has developed rapidly in recent years and achieved excellent results on various benchmark datasets. However, these datasets are usually composed of high-quality [high resolution (HR), high signal-to-noise ratio, etc.] images. In many scenarios, we need to detect objects in low-resolution (LR) images. But the detector trained in HR images performs poorly in LR images. The image super-resolution (SR) algorithm is an important image processing technology that has been proven to improve the performance of various visual tasks. Based on this consideration, we combine image SR technology with the object detection task to design an object detection algorithm for detecting LR images. Specifically, we propose a lightweight SR algorithm that achieves a good balance between parameters and performance. We perform SR reconstruction on the LR image in advance and detect the reconstructed image instead of directly detecting the LR image. Both quantitative and qualitative experimental results show that our LR object detection algorithm significantly improves the detection performance of LR images.

Many algorithms to enhance low-light images have been proposed. They often assume that low-light images are clean. However, images acquired in low-light environments often contain noise. Consequently, those enhancement algorithms often simultaneously amplify the noise. For the purpose of overcoming the above-mentioned shortcoming, we propose a framework of joint enhancement and denoising. It is accomplished using the non-subsampled shearlet transform with adaptive soft-threshold shrinkage. The multi-scale feature of the proposed model can better represent edge textures in an image and decomposes an image into two parts of low frequency and high frequency for separate processing. Experimental results show that, compared with state-of-the-art methods, the proposed algorithm can achieve better quality in terms of objective assessment and subjective assessment.

Point cloud registration is a crucial task in fields such as three-dimensional reconstruction, target localization, and simultaneous localization and mapping. In the case of relatively low overlap, points out of the overlapping region act as interferences that will have a terrible impact on the registration results. Some existing registration methods improved for low-overlap situations still suffer from poor accuracy and low success rate. Therefore, a reliable two-stage registration method is introduced. First, a quality first sample consensus algorithm is used to determine the initial points for optimal transport and complete the coarse registration. Second, a reliable optimal transport algorithm is proposed for fine registration, in which the points for transmission are dynamically adjusted according to the iteration of the transport plan to improve the registration efficiency. In the registration experiments with the standard models and real scene models, this method reached 1e−02 and 1e−01 orders of magnitude, respectively, which outperforms better than the existing mainstream algorithms. This method can still perform excellent and stable registration results on multiple models and various missing cases, and the success rate is kept at more than 90%.

Current learning-based dehazing methods simply rely on the paired synthetic datasets and physical models, which can hardly describe the complicated degradation for practical applications. These methods still struggle to achieve haze removal, color distortion, and detail restoration instantly, and they ignore the frequency characteristic differences and prior knowledge importance. To address these problems, we propose an unpaired stage-wise framework integrating frequency-guided filtering and progressive physics learning in an adversarial dehazing network, called FPD-Net. To be specific, a guided filter based on frequency information is employed to decompose the high and low frequency components for better feature extraction. We further merge the prior and physical knowledge to form progressive physics learning, which produces pleasing haze-free outputs with high visibility and reality. For better atmospheric light estimation, the variational auto-encoder and Kullback–Leibler loss are included to represent the illumination message. Extensive experiments on both synthetic and real datasets prove that our designed FPD-Net achieves better performance visually and quantitatively than the comparison dehazing models.

Generative steganography is a research hotspot particularly with regard to information hiding, which involves the hiding of secret information by generating sufficiently “real” secret media. In recent years, generative steganography schemes have made significant progress in images, but the field of video steganography is still in the exploratory stage. Combined with deep convolutional generative adversarial nets (DCGAN), a semi-generative video steganography scheme based on a digital Carden Grille was proposed. A dual-stream video generation network based on DCGAN was designed to generate three parts of videos: foreground, background, and mask; generation network produced different videos with random noise. The digital Carden Grille is used as the key for embedding and extraction. The sender can generate a digital Carden Grille in the mask through two different methods, reasonably assign the embedding capacity among RGB channels, and use video pixels as the carrier to achieve the information embedding in a semi-generative way. The receiver can determine the embedding position of information through the Carden Grille and extract the secret information hidden in the pixels. Experimental results show that the stego video generated by this scheme has good visual quality, with a Frechet inception distance score of 92. The embedded capacity is better than the existing generative steganography schemes, up to 0.12 bpp. Using Syndrome Trellis Coding, the proposed scheme can transmit secret messages more efficiently and securely.

Traditional simultaneous localization and mapping (SLAM) algorithms often rely on developers manually setting fixed parameter thresholds based on their experience. However, as the application scenario changes, these parameters often require manual readjustment. To address this limitation, we proposed a method that leverages the deep deterministic policy gradient (DDPG) to adjust visual SLAM parameters adaptively. This method involves selecting visual SLAM parameters to create a continuous action space. We have used the uncertainty of pose to construct a reward function. The agent acts on the visual SLAM environment by selecting the best action based on the output of the DDPG network, and the network parameters are updated according to the environmental feedback. We have proposed a DDPG-based network with improved performance by incorporating convolutional networks and gate recurrent unit. Finally, the experiments on the Euroc MAV and TUM-VI datasets have been conducted. The results demonstrated that the method significantly improved the accuracy of pose trajectory in indoor scenarios while avoiding the cumbersome parameter adjustment process.

The smart walker is a special service robot that can assist the elderly in walking. Deploying gait recognition algorithms on the smart walker can help the elderly identify relatives and friends from a distance in real time and obtain some pedestrian information to avoid danger. However, in actual usage scenarios, it is difficult to quickly obtain qualified pedestrian silhouettes as input images for gait recognition algorithms. The difficulty is that unlike fixed cameras in traditional gait recognition scenarios, the camera of a smart walker acquires a dynamic background due to the movement of the robot, and it is difficult to accurately segment dynamic pedestrians from the dynamic background. Instance segmentation algorithms commonly used to segment silhouettes in static images require high-resolution input images. To quickly segment gait silhouettes from dynamic backgrounds, we propose a color filter-based gait silhouette extraction method (CFSE) in dynamic background. We use instance segmentation algorithms to distinguish the foreground and background, use background subtraction algorithms to obtain portrait silhouettes, and finally use color filtering to eliminate background colors to get the silhouette. Experimental results show that the method we proposed can extract silhouettes more accurately in moving scenes, and the algorithm speed meets the real-time requirements.

The shape and values of a typical static convolution kernel remain fixed once the network is trained. Recently, dynamic convolutions were proposed to change the kernel’s values depending on the input during the test phase. We aim to extend the concept of dynamic convolutions by introducing an element-wise dynamic convolution approach. This method enables adaptive changes in kernel values for each output data element. Furthermore, a deformable element-wise dynamic convolution is proposed to enable simultaneous changes in kernel shape and value. The proposed deformable dynamic convolution is compatible with the static convolution in terms of input–output relationships. The capability of existing network architectures can be enhanced by replacing the static convolution with the suggested deformable dynamic convolution. Extensive experiments demonstrate that the proposed deformable dynamic convolution can improve the network performance in various computer vision tasks, including image classification and semantic segmentation.

In practical applications, to improve the real-time performance of end-to-end stereo matching networks, the existing methods build cost volume at low resolution. However, with detailed information missing in low-resolution features, it is difficult to get accurate disparity estimation results in weak texture regions. Besides, smooth L1 loss supervision also results in a loss of accuracy in disparity discontinuity areas. To solve these problems, we propose an efficient stereo-matching network based on multiple attention mechanisms and edge optimization, which can achieve high accuracy in a short time. The multi-scale attention module is applied to enhance the feature expression in detail regions. For weak texture areas, we construct a concatenation cost volume and a multi-level patch matching volume, which can be combined to improve the network’s attention to weak texture regions. In terms of edge optimization, we perform bimodal Laplace modeling of the sampled edge points’ disparity distribution and optimize the edge region of the initial disparity map using likelihood loss to obtain sharp edges. The experimental results show that, on the SceneFlow and KITTI datasets, the proposed network improves by 32% and 27% in accuracy compared with BGNet+.

Large-scale point cloud semantic segmentation is crucial for understanding the scene environment in a variety of applications such as autonomous driving. To address the problem that existing large-scale semantic segmentation methods cannot fully exploit the local neighborhood information of point clouds and are prone point features, we propose a local parallel feature extraction (LPFE) module that enhances the feature representation of points by separately coding the semantic information and geometric structure of each point in point clouds. The LPFE module consists of three blocks, including the semantic information encoding (SIE) block, the position information encoding (PIE) block, and the parallel mixed pooling aggregation (PMPA) block. (1) The SIE block uses the proposed adaptive feature fusion strategy to generate rich and higher-level semantic features for each point. (2) The PIE block enables the network to extract fine-grained geometric features while focusing on the entire local region by establishing relative and global position relationships in the neighborhood. (3) The PMPA block uses a bilateral structure to capture both the maximum local features and local contextual features using mixed pooling operations. The above blocks form a new network for semantic segmentation of large-scale point clouds, called “local parallel feature extraction network,” which achieves competitive performance for semantic segmentation on large-scale point cloud datasets S3DIS and Semantic3D, and demonstrates the effectiveness of the modules through specific ablation experiments.

Accurate detection of flames is an important prerequisite for the implementation of many security measures. To address the problems of low detection accuracy and difficulty in detecting small target flame in the current mainstream flame detection algorithms, we propose a small target flame detection algorithm incorporating an attention mechanism. First, we optimize the self-acquired dataset with data enhancement and structural similarity algorithms to improve the quality of the dataset and enhance the training efficiency of the model; then, a coordinate attention mechanism is introduced in the feature extraction stage of the original you-only-look-once (YOLOv7) network to improve the ability of the model to extract useful feature information and help the model to more accurately; and then, a small target flame detection layer is introduced at the output of the YOLOv7 network to improve the model’s ability to detect small flames. Finally, the parameters of each layer of the improved model are initialized through model pre-training and migration learning to reduce the risk of model overfitting and accelerate the convergence speed of model training. The experimental results show that the average detection accuracy of the flame detection algorithm designed in this paper reaches up to 89.5%, which is 8.7% better than the detection accuracy of the original YOLOv7 algorithm and is innovative and practical.

Recent developments in advanced generative deep learning techniques have led to considerable progress in deepfake technology. CNN-based deepfake detection approaches have demonstrated superior performance. The ability to learn meaningful representations generated by convolutional multilayer nonlinear structures is the key to success. However, the black-box nature of such approaches has been a major concern for exploring hidden and complex characteristics as well as potential limitations of CNN-based models. To gain insights into the scope of the deepfake detection task, we investigate the effectiveness of handcrafted feature-based methods for deepfake video detection. First, we experiment with six top-performing handcrafted descriptors to extract the discriminating image features and then train SVMs on the extracted features to learn a suitable model. We also study the effect of selecting specific facial components on the detection performance. Specifically, we consider features extracted from the left eye, right eye, mouth, and entire face. Moreover, we propose a combination of these features and highlight the importance of this combination in terms of detection performance. Experimental results show that the SIFT feature descriptor achieves the best performance on deepfake videos generated by the neural texture technique, with a detection accuracy of 83.50%, which is better than deep learning-based methods. This is in contrast to the conventional understanding that deep learning methods systematically outperform handcrafted feature-based approaches. In addition, the obtained results on the FaceForensics++ dataset highlight the benefit of using some facial components to further boost the detection performance. Moreover, motivated by the effectiveness of the LBPTOP and SIFT in the deepfake detection task, we combined the LBPTOP and SIFT to best characterize the specific spatiotemporal inconsistencies commonly found in fake videos for boosting deepfake detection performance. Finally, we show the strengths and weaknesses of methods based on handcrafted features for deepfake detection and provide directions for future research.

We propose a novel image encryption scheme that utilizes chaotic maps and random scan patterns to protect image content. Our scheme includes three stages: pixel permutation via random scan patterns, block pixel diffusion, and block permutation. In the first stage, the input image is partitioned into overlapping blocks, and random scan patterns controlled by chaotic maps are exploited to scramble pixels of blocks. In the second stage, content-based keys are generated from the input image, and the content-based keys and chaotic maps are used to calculate secure matrices for achieving diffusion. The block pixel diffusion is conducted by performing element-wise XOR operation between blocks and their corresponding secure matrices. In the third stage, the image is partitioned into non-overlapping blocks, and all blocks are randomly permuted to further improve security. Key space analysis shows that the key space of the proposed scheme is bigger than those of some existing schemes. Experiments on standard test images are conducted by using some well-known metrics, such as histogram, χ² test, correlation coefficient, entropy, the number of pixels change rate, the unified average changing intensity, and peak signal-to-noise ratio. The results demonstrate that the proposed scheme reaches the state-of-the-art performance in security and speed.

Multi-view subspace clustering has attracted widespread attention for its superior clustering effectiveness. However, most of the current methods do not fully exploit the intra-view (sample-to-sample) and inter-view (view-to-view) information among multi-view data. Focusing on this problem, we propose a method called enhanced intra-inter view correlation learning for multi-view subspace clustering (EIVCL). EIVCL simultaneously considers the intra-view and inter-view correlations and introduces two constraint terms for each aspect to capture the data structure information. Specifically, in the intra-view space, we apply the tensor-singular value decomposition (t-SVD)-based tensor nuclear norm (TNN) on the tensor formed by stacking self-representative coefficient matrices for obtaining the high-order correlation information between samples in the specific view. Moreover, we introduce a hypergraph-induced Laplace regularization term to preserve the local geometric structure within views. In the inter-view space, we impose the t-SVD-based TNN on the rotated tensor to obtain the multiple views correlation. Furthermore, we utilize the kernel dependence metric, namely Hilbert–Schmidt independence criterion, to capture the high-order non-linear relationships between views. In addition, all above strategies are integrated into a unified clustering framework, which is solved by our proposed optimization algorithm based on the alternating direction method of multipliers. Extensive experiments on six benchmark datasets demonstrate that EIVCL outperforms several state-of-the-art multi-view algorithms.

Few-shot image classification learns from the limited samples of known categories, with the purpose of identifying new categories of samples. However, due to the under-representation of the class prototype generated by support samples for few-shot image classification, it leads to large classification errors. Therefore, we propose the multi-loss dual prototype network. First, the class prototype of the support samples and the class prototype of the query samples are constructed by the support samples or query samples, respectively. Second, the local loss of self-calibration (SC) is obtained by calculating the difference between the class prototype features of the support samples and the support sample features. The local loss of mutual-calibration (MC) is obtained by computing the difference between the class prototype features of the query sample and the support sample features, as well as the difference between the class prototype features of the support samples and the query sample features. Then, the global loss of the support samples and the query sample classification is calculated, and the local and global losses are weighted as the total loss to train the network. The total loss includes both SC and MC, which enables the network to make full use of the sample feature information. To adaptively enhance the features of the target region, we propose a cross-channel spatial attention model that generates class prototype features that are more consistent with the feature distribution of the class. Finally, we improve the transductive inference algorithm by assigning weights to the unlabeled query samples using cosine similarity and expanding the weighted query samples into the support set through iteration to further enrich the class prototype features. Experimental results show that the proposed method is effective.

Few-shot learning consists in training a classifier, which can be quickly adapted to new tasks with a few samples. To address the few-shot learning tasks, metric-based meta-learning methods explore the appropriate metrices to measure the similarity between the support and the query samples. However, existing methods ignore the similarity relationship between the labeled samples in the support set. Consequently, we propose a contrastive knowledge-augmented self-distillation approach to leverage the similarity relationship between a few labeled samples in the support set and allow the model to focus on more regions of images. Specifically, we calculate the classification probability of query images and each class prototype, respectively, and consider the classification probability of each class prototype as the teacher to guide the classification of query samples. Meanwhile, we design a contrast loss to bring the eigenvectors of the same class closer and push the eigenvectors of different classes further. In addition, a transformation function is implemented to allow the model to focus on more regions of the images, so as to obtain the key features. Extensive experiments are conducted on miniImageNet, tieredImageNet, and Caltech-UCSD birds 200, and the results show that our method can enhance metric-based meta-learning methods and outperforms the state-of-the-art methods.

Most trackers based on the correlation operation mainly carry out a straightforward fusion of template features and search region features. The correlation operation, however, is a local linear matching process that might cause the loss of semantic information, leading to local optimums. We propose a tracker method that takes advantage of both the attention mechanism and feature fusion. This approach uses an attention-based feature fusion network to integrate the characteristics of the template area and search area, which effectively avoids the semantic information loss and local optimum induced by the correlation operation. Meanwhile, by adopting a modified ResNet50 network for extracting multilayer target features to fully utilize both shallow and deep features, our method takes full advantage of the template information to get better tracking results. Extensive experiments on four large-scale tracking benchmark datasets—VOT2019, OTB100, GOT-10k, and LaSOT—show that our tracker obtains more accurate tracking results and has a better performance.

Three-dimensional (3D) object detection is crucial for accurate recognition of autonomous driving roads, and the distribution of point clouds in 3D scenes becomes sparse with increasing distance, thus seriously affecting the sensor’s perception precision. To address this problem, we propose a two-stage 3D object detection network based on point and voxel feature fusion. In the first stage, a spatial semantic feature fusion module is designed to effectively fuse low-level spatial features and high-level semantic features to generate high-quality proposals. Then, an attention mechanism-based residual module is constructed to expand the receptive field and adaptively aggregate the voxel features in the 3D scene. At the same time, the sampled key points and voxel features are fused to extract the key information in the 3D scene. In the second stage, the graph network pooling module is introduced to construct local graphs on 3D proposals using key point features as nodes to estimate the confidence and location of objects more accurately. Experimental results on the KITTI dataset show that the detection precision is improved significantly in easy, moderate, and hard tasks.

Plant diseases are a major concern for farmers worldwide as they can cause significant economic losses. Among vegetable crops, cucumber is particularly susceptible to various diseases, including fusarium and mildew. We present a unique dataset of cucumber leaf images collected from the Carsamba and Silifke districts of Turkey. This dataset is used in both newly proposed convolutional neural networks and pre-trained models available in the literature in addition to new ensemble models for disease prediction and classification. To optimize the classification accuracy, the grid search method is employed to identify the best combinations of base models and ensemble learning methods. The classification process is carried out using newly designed triple ensemble models and utilized fivefold cross-validation, and the results are evaluated using accuracy, F1-score, and receiver operating characteristic curves. The highest accuracy rate achieved in our study is 93.79%, which demonstrates the effectiveness of our approach for accurately identifying cucumber leaf diseases. Our proposed approach is not only accurate but also effective, making it a practical solution for farmers to use to quickly identify and control diseases in their crops. Our study provides valuable insights into the application of deep learning techniques for the classification of cucumber plant diseases and can serve as a useful model for further research in this field.