Accurate and efficient corrosion detection is a difficult but important issue with immediate relevance to maintenance of Naval ships. The current process requires an inspector to physically access the space and perform a very manual visual inspection of the space. Considering the schedules of both the inspector and the ship, coordinating the inspection of hundreds of tanks and voids is not always a straightforward process. There is a significant amount of research into automatic detection of corrosion via computer vision algorithms, but performing pixel level segmentation introduces added difficulty. There are two key reasons for this: the lack of annotated data and the inherent difficulty in the type of problem. In this work, we utilized a combination of annotated data from a different domain and a small hand labeled dataset of panoramic images from our target domain: the inside of empty ship tanks and voids. We trained two High-Resolution Network (HRNet) models for our corrosion detector; the first with a dataset outside our target domain, the second with our hand annotated panoramic tank images. By ensembling our two models, the F1-score increased by about 120% and IOU score by about 176% with respect to the single baseline corrosion detector. The data collection process via LiDAR scanning allows the inspection process to be performed remotely. Additionally, the setup of the detector leads to a natural expansion of the corrosion dataset as panoramas from LiDAR scans are continually fed through the detector and the detections are validated. This allows for the corrosion models to be later retrained for potential improvement in accuracy and robustness.
Adversarial training (AT) is considered the most effective strategy to defend a machine learning model against adversarial attacks. There are many different methods to perform AT, but the underlying principle is the same, namely, augment the training data with adversarial examples. In this work, we investigate the efficacy of four different adversarial example generation strategies on AT of a given classification model. The four methods represent different categories of attack and data. Specifically, two of the adversarial generation algorithms perform attacks in the pixel domain, while others operate in the latent space of the data. On the other hand, two of the methods generate adversarial data samples designed to be near the model decision boundaries, while the other two generate generic adversarial examples (not necessarily at the boundary). The adversarial examples from these methods are used to adversarially train models on MNIST and CIFAR10. In the absence of a good metric to measure robustness of a model, capturing the effect of AT using a single metric can be a challenge. Hence, we opt to evaluate the robustness improvements resulting of the adversarially trained model using a variety of empirical metrics introduced in the literature that measure local Lipshitz value of a network (CLEVER), smoothness of decision boundaries, robustness to adversarial perturbations and defense transferability
KEYWORDS: 3D modeling, Data modeling, Image quality, Image classification, RGB color model, Light sources and illumination, 3D image processing, Satellite imaging, Earth observing sensors, Satellites
Synthetic data has shown to be an effective proxy for real data in order to train computer vision algorithms when acquiring labeled data is costly or impossible. Ship detection and classification from satellite imagery and surveillance video is one such area, and images generated using gaming engines such as Unity3D have been used successfully to circumvent the need for annotated real data. However, there is a lack of understanding of the effect of rendering quality of 3D models on algorithms that use synthetic data. In this work, we investigate how the level of detail (LOD) of objects in a maritime scene affects ship classification algorithms. To study this systematically, we create datasets featuring objects with varying LODs and observe their significance in computer vision algorithms. Specifically, we evaluate the impact of mismatched LOD datasets on classification algorithms, and investigate the effect of low or high LOD datasets on a model's ability to transfer to real data. The LOD of 3D objects are quantified using image quality metrics while the performance of computer vision algorithms is compared using accuracy metrics.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.