“This PDF file contains the front matter associated with SPIE Proceedings Volume 12783, including the Title Page, Copyright information, Table of Contents, Committee Page and Introduction.

Research into new functional materials has been ongoing on for many years. The successes are based on a classic trialand-error method. In the years that followed, various methods such as computer-aided calculations and high-throughput screening were added. Since the beginning of the 21st century, immense progress has been made in the field of artificial intelligence, which has since found its way into a wide variety of specialist disciplines and everyday life. With the advent of artificial intelligence in the research of new materials, there is hope for new results and savings in time and money. The approach presented here serves to monitor crystallization processes. Crystallization processes are used to evaporate new compositions of substances dissolved in a solvent. Evaporation produces crystals, which are then used for further investigations into the material properties. However, the crystallization process is very time-consuming and highly dependent on the solution and the environmental parameters. As a result, the timing of the process is difficult to predict and very lengthy. Therefore, this paper presents a method combines two areas, computer vision and artificial intelligence, and thus offers the possibility to monitor a crystallization process. The significant points, the start and end point, are detected, and the course of the crystallization process over time is also recorded. For this purpose, a pre-trained ResNet34 network is used, which has been trained on the characteristics of crystals through transfer learning, and a visual analyzer unit for in-situ sample acquisition. With this precise measurement setup, crystallization processes can be monitored and subsequently automated. This can save time and money and accelerate research into new materials.

The solutions obtained by training the deep neural network are highly dependent on the parameters including the learning rate. Therefore, finding the appropriate settings for training deep neural networks is very important. In particular, it is necessary to find the better settings for SOTA models of Vision Transformer(ViT), whose structure is different from ordinal models. In this paper, we focus on the learning rate to find a better value using the Learning Rate Range Test (LRRT). Through our experiments, we found that the appropriate LR is located where the decrease in loss value stops in the LRRT. In addition, we discuss about the effects of the number of epochs and the LR warm up.

The research on CNN applications for medical image processing has been progressing rapidly. In the process of the development, hurdles appear and are to be overcome. The limitation in training samples is one of them, and restriction in computation resources can be another. In this paper, we present a design approach of application specific CNN (ASCNN), allowing to minimize the computational complexity of CNN systems without lowering the performance. This approach is to full-custom design CNNs for specific applications, such as brain tumor detection, so that each part of a CNN can be optimized to suit the input data and the task assigned to it. The convolution kernels and layers are made just-sufficient, nothing excessive. In this way, the randomness and the redundancy in computation can be minimized, the dependency on training samples decreased, the information density in data flow increased, the computation efficiency/quality and performance reliability improved. Three ASCNN systems for brain tumor detection are also presented as design examples. The results of the performance evaluation demonstrate that each of them delivers a high-quality detection with a computation volume of one-digit percentage, or less, of that needed by other CNN systems recently reported in reputed journals in the research area. Hence, ASCNN approach is effective to achieve high process quality at low computation cost. It can also lower the barrier of resource requirement of CNN systems to make them more implementable and applicable for general public.

In this paper, we propose a high-resolution GAN model for image dehazing in icing meteorological environment, which strictly follows a physics-driven scattering strategy. First of all, the utilization of sub-pixel convolution realizes the model to remove image artifacts and generate high-resolution images. Secondly, we use Patch-GAN in the discriminator to drive the generator to generate a haze-free image by capturing the details and local information of the image. Furthermore, to restore the texture information of the hazy image and reduce color distortion, the model is jointly trained by multiple loss functions. Experiments show the proposed method achieves advanced performance for image dehazing in icing weather environment.

Trajectory compression can solve the redundancy problem of a large amount of trajectory data generated by GPS positioning systems. This paper proposes to find the feature points inside the trajectory according to the motion trend and variable speed characteristics, and then use these feature points to segment the original trajectory and compress them separately. Experiments show that the algorithm performs well in terms of running time, compression rate and average error.

Object detection plays an important role in autonomous driving. In the past decades, many object detection methods relied on 2D images, losing spatial information due to projecting 3D space into 2D space. Recently, LiDAR has become a popular sensor for 3D point cloud target detection. This paper proposes a new RCNN detection framework based on pseudo segmentation (PS-RCNN). This model is designed to achieve accurate and efficient detection on point cloud by transmitting feature information reversely. The information transmission is supervised by the semantic segmentation task. In order to reduce the difficulty in labeling, a novel algorithm is designed to generate segmentation pseudo-labels. Experimental results conducted on KITTI Dataset and Waymo Open Dataset demonstrate that our model outperforms its counterparts for detecting small objects with a balance between accuracy and efficiency.

Compressed sensing is a signal processing technique used for signal reconstruction with significantly smaller number of samples than the requirements of the Nyquist-Shannon theorem. In this work, we simulate a lensless digital holographic system. We investigate the ringing-like artefact introduced by truncation by the camera aperture. We present the results of using the orthogonal matching pursuit based compressed sensing algorithms to combat this ringing-like artefact. We demonstrate that compressed sensing achieves remarkable reconstructions and suppresses ringing well, but only up to a point in terms of the size of the aperture. This research could help the advancement of compressive digital holography.

Nowadays, dysmorphic disorder among contemporary adolescents has attracted more and more attention from people of all social circles. The purpose of this study is to provide a useful self-evaluation model of adolescent image for assessing adolescents’ dysmorphic disorder situations. 249 teenagers participated in this study and various machine learning algorithms have been developed and utilized for building the self-evaluation model, such as the K-Nearest Neighbor algorithm, Naïve Bayes algorithm, and Principal Component Analysis algorithm. The best self-evaluation model developed in this project gave the highest accuracy of 76.92% on the testing set. For predicting the trend of dysmorphic disorder among contemporary Chinese adolescents, ordinary least squares linear regression model has been created, and then the percentages of different age stages to carry out major plastic surgery in 2022, 2023, and 2024 have been predicted

With the development of monitoring technology and the improvement of people's security awareness, intelligent human abnormal action recognition technology in the field of action recognition is increasingly high. In most cases, abnormal human action may have little difference in appearance compared with normal behavior, so the control of visual rhythm information becomes an important factor affecting action recognition, but people often focus on the appearance information of the action and ignore the rhythm information. In this paper, we introduce the temporal pyramid module to process the visual tempos information, meanwhile, the traditional LSTM local history information transfer method is very easy to lose the context information, which is not conducive to the grasp of global information and thus will greatly affect the processing effect of the temporal pyramid. This paper introduces a non-local neural network module to enhance the network's ability to grasp global information and the model's long-range modeling capability, which is used to supplement the temporal pyramid module. Finally, this paper uses the mainstream anomaly dataset UCF-Crime to test the network performance, and the improved network model recognition accuracy AUC reaches 0.82, which is better than other stateof-the-art methods.

The scale of the monitoring system is becoming larger and larger. In order to perform intelligent video processing in surveillance systems, we need to detect moving object in image sequence. Some methods in the literature can achieve a valid detection result, but usually they have high computational complexity. In the outdoor scenes, the background is usually dynamic, and the dynamic background makes it difficult to detect moving object. In order to solve these problems, we propose a new method with low computational complexity using mass center coordinate to expand the mask image. The proposed method can remove the interference of dynamic background in the detection. Experiment results show that our method can mask the dynamic background more completely while ensuring fast computation and consuming less hardware resources. The method can be used in massive video intelligent processing.

Car owners altering license plates using different typefaces and designs violate the law that strictly forbids such behaviour. Traffic police officers claim that changing the license plates makes it impossible to read the registration numbers due to an increase in fatal street collisions and car thefts. They worry that it may be difficult to track down vehicles used in hit-and-run incidents. It is challenging to impose further limitations on any algorithm used to identify and recognise license plates in a developing nation like Bangladesh. This work has the primary objective of designing a reliable detection and recognition system for transitional, standard car license plates, which are frequently seen in developing countries. Increase the effectiveness of reading license plates drawn or printed in various styles and typefaces employing cuttingedge technology, including machine learning (ML) models. For this study, You Only Look Once (YOLOv3) is used to utilising the most recent version of the object detection method. The raw image is pre-processed to increase its quality and then divided into appropriate-sized grid cells to determine where the license plate should be placed. After that, the data is post-processed, and the accuracy of the proposed model is evaluted using industry-recognised standards. A sizeable image dataset was used to be tested using this proposed methodology. The presented system is expected to be essential for vehicle monitoring, parking fee collection, lowering traffic accidents, and identifying unregistered vehicles. The results demonstrate that the suggested method achieves 97.1% mAP, 95.3% precision and 96.8% in plate detection

Post Pandemic world of Covid-19 has set human race to a transitioned frequency. Through this transitional period the world needs to serve itself with much needed technology and services. The need of time is now to build the system compatible to the crisis we face together that ensures a risk-free safe environment. New regulations and measures have been established in order to provide safety that includes regular wear of face mask. It is necessary to strictly execute this new rule in public places that helps reducing the spread of virus. Along the mask, a mask detector system comes hand in hand playing a vital role. On official image recording sectors, the mask detection is a requirement to ensure the identity of a personnel. This surveillance method enables the alert system to remove the mask while taking any photograph for documentation that demands clear photographic identity. In this paper, we have briefly discussed this approach of face mask detection system while capturing a photo using deep learning algorithms.

Increasing security awareness in the public sector are leading to a more and more widespread use of surveillance applications. Although the available technologies like video processing are already well advanced, they still suffer from high false alarm rates when used under realistic conditions. We present a method for sensor fusion based on probability density maps and a rule engine. The system was tested in a public area using the combination of audio localization, audio classification and video detection using 79 simulated scenarios and 44 hours of sample data recorded over a period of several weeks. The false positive rate decreased by 60% and the event localization rate increased by 25% with the fusion approach compared to the detection performance of individual techniques

At present, the workload of mental workers in society is getting heavier and heavier, and it is necessary to assess their health status. Compared with other physiological signals, the pulse is easy to obtain and non-invasive. In this paper, through pulse signal detection, pulse data preprocessing and feature extraction, 12 sets of feature values are selected. Then based on these feature data, using support vector machine algorithm modeling, for different testers to build different personalized human physiological state discrimination system. The experimental results show that the classification accuracy rate reaches 91.17%, which proves that the selected feature value has a strong correlation with the physiological state, and the classifier is effective.

Traditional direction-of-arrival (DOA) estimation methods use a single processor to deal with the array data. In recent years, the increasing of the scale of sensor arrays brings heavy workload for single processor. Distributed optimization based on multiple local processors has become one of the current research hotspots due to the advantage of parallel computing. In this paper, we proposed a distributed DOA estimation method for massive large-scale arrays. First of all, we provide the signal model and the distributed optimization problem based on sparse representation in a distributed framework. Then, the optimization problem is solved by the alternating direction multiplier method (ADMM), where the overall structure of array is not changed. Compared with the centralized method, our distributed method can greatly reduce the computational complexity while ensuring the estimation accuracy under the large aperture array. Simulation results are provided to show the superiorities of our method.

In this paper, a magnetic-force based thickness measurement model was established based on the physical properties of radar absorbing coatings that are easy to be magnetized and the experimental data obtained from the standard samples with titanium alloy substrate coatings. The experimental data obtained from the thickness measurement experiment of the standard samples with titanium alloy substrate coatings were fitted by using least squares method. A thickness measurement sensor was developed based on the STM32 microcontroller, which solved the problem that the traditional wireless design was unable to balance transmission distance, anti-interference, and power consumption. On this basis, thickness measurement software was developed based on WinCE 6.0 and measurement experiments were done. The experimental results showed that this magnetic-force based thick measurement system had the advantages of simple structure and high software reliability.

Wireless sensor network is domain-specific wireless system composed of various nodes called sensor nodes. The purpose of the nodes is to collect physical and environmental data parameters such as sound, pressure and temperature, and relay the data across the network through a central node. The goal is to get the data from the sensors in a secure manner. Data flow in wireless sensor networks is threatened by a variety of attack types, including Sybil, Wormhole, Sinkhole, and others. In order to accomplish privacy, integrity, availability, confidentiality, a number of protocols and techniques have been developed. In this paper we will provide a systematic review that spots the light on the common threats and the countermeasures proposed by the previous studies

Unlike other stock markets participants, the participants in China mainland are composed of individual investors, which account for 82% of the trading volume of the stock market. The decision-making basis of individual investors is mainly public opinion and recent stock prices. Therefore, the public opinion on professional stock social sites has an important impact on the decision of individual investors, which in turn affects the trend of the stock market. However, the previous stock market forecasting methods mostly ignored the influence of public opinion information on the market. For this reason, this paper proposes a novel framework to predict the stock trend by using both public opinion and stock numerical data. The original contributions of this paper include stock commentary word embedding model based on the stock comment text data crawled from https://xueqiu.com through two-stage training and LSTM-CNN layered model based on the improved self-attention mechanism. Two main experiments are conducted: the first experiment extract stock commentary word embedding, and the second experiment forecasts the stock price trends of Shanghai and Shenzhen A-share market. Results show that: 1)LSTM-CNN layered model is better than previous methods; 2)The combination of public opinion information and numerical data can improve the performance of the model; 3)Stock commentary word embedding model is better than pre-training word embedding model; 4) The longer the data span, the better the stock forecasting model will perform

In this paper, the on-board computer software and its reliability design of a micro-satellite are analyzed. Because the satellites can’t be repaired while flying in the space, it should be as reliable as possible. The micro-satellite on-board computer adopts the dual-computer redundancy method, and uses the dual-computer communication to realize the system operation state recovery and hold when the host and standby computer switch, and some reliability design of the computer hardware in the space environment is realized and completed by the software. The space flight test proves that the micro-satellite on-board computer software is safe and reliable. This paper introduces the main features of the microsatellite's On-board Data Handling software and the reliability design technology adopted, which reflects the microsatellite's On-Board Data Handling software's main design technology in terms of reliability.

Research on rapid clustering technology based on bibliographic information of scientific and technical literature aims to efficiently realize the correlation analysis of scientific and technical literature, laying the foundation for discovering hot spots and trends in the research field, conducting interdisciplinary and cross-border research, and accurately recommending scientific and technical literature. Focusing on the analysis of clustering algorithms, we proposed an improved k-meansbased Firefly Algorithm k-means (FAKM) clustering method, which effectively solved the problem of randomly selecting the initial center points of class cluster when using k-means algorithm for clustering in the clustering stage, which leads to local optimum, low accuracy and large gap between the division of class clusters and the real situation of clustering results. The use of FAKM clustering algorithm resulted in better clustering performance, high accuracy, and fewer iterations. The experimental results showed that the method achieved a silhouette coefficient of 0.54 and adjust mutual information of 0.69 on the same scientific and technical literature data set, which proved the good performance of the method.

Dataspace is a new way of data integration. Entity resolution identifies two records that point to the same entity in the real world. In this paper, a record graph is constructed by using the records in the data set. The redundant comparisons are removed by pruning the record graph, and the records is divided into blocks according to the pruned graph. The subsequent entity resolution work is only carried out in blocks. When the entity is parsed in the block, the method of attribute mapping and expression representing attribute value is used to further divide the data to ensure the accuracy of parsing. Methods experiments were carried out on real data sets.

While Bitcoin has been a hot topic in the investment world due to its rising value, gold remains a popular investment option. To create a value prediction model for the best investment strategy, we utilized LSTM and found that it had a higher fitting effect than other two models, grey prediction and time series. The accuracy rate is 88.7%, the loss rate is 0.135%. We test different batch sizes to ensure the accuracy of prediction and established an appropriate algorithm to calculate the best investment strategy for each day. To test the accuracy of the model, we use four methods, including testing the accuracy of the risk factor in the model, observing the growth of total asset value, calculating the error rate of the investment process, and performing robustness analysis under low investment costs. We also perform sensitivity analysis to determine the impact of transaction costs on the strategy and results. The results show that the fluctuation in Bitcoin transaction costs is more significant and can affect the frequency of trading activities, ultimately affecting the final profit.

This paper develops a model for predicting Kweichow Moutai stock prices combining both fundamental analysis (Composite Index, earnings per share, Market capitalisation in circulation) and technical analysis (past stock prices) based on the transformer deep learning method. The approach improves the accuracy of stock price prediction that relies solely on fundamental analysis or technical analysis.