21 February 2025 in IEEE Access
Surface defects, including cracks, scratches, and deformations, significantly affect the product performance and service life in industrial manufacturing. Accurate detection of metal surface defects ensures product quality and reliability. Traditional nondestructive testing (NDT) methods, such as ultrasonic and eddy current inspections, are effective but face challenges in terms of efficiency and scalability for industrial applications. This review examines the recent advancements in automated defect detection, focusing on image acquisition, image processing, and detection techniques. The study also evaluates state-of-the-art machine learning and deep learning methods used in defect detection, including the symmetric convolutional network (SCN), swin transformer you only look once (ST-YOLO), and statistical texture feature enhancement network (STFE-Net), which achieve high accuracy in defect classification and detection. In addition, advanced three-dimensional (3D) detection methods such as photometric stereo, light-field imaging, and structured light have been explored for their potential to address challenges such as real-time detection, few-shot detection, and small-target detection. The study concludes by identifying current limitations and proposing future directions for enhancing industrial defect detection systems.
21 October 2019 in IEEE Access
Inspired by their tremendous success in optical image detection and classification, convolutional neural networks (CNNs) have recently been used in synthetic aperture radar automatic target recognition (SAR-ATR). Although CNN-based methods can achieve excellent recognition performance, it is difficult to collect a large number of real SAR images available for training. In this paper, we introduce simulated SAR data to alleviate the problem of insufficient training data. To address domain shift and task transfer problems caused by differences between simulated and real data, we propose a model that integrates meta-learning and adversarial domain adaptation. We use sufficient simulated data and a few real data to pretrain the model. After fine-tuning, the pre-trained model can quickly adapt to new tasks in real data. Extensive experimental results obtained in the real SAR dataset demonstrate that our model effectively solves the crossdomain and cross-task transfer problem. Compared with conventional SAR-ATR methods, the proposed model can achieve better recognition performance with a small amount of training data.
18 December 2018 in IEEE Geoscience and Remote Sensing Letters
A variety of machine learning approaches have been applied to synthetic aperture radar (SAR) automatic target recognition. The performances of these approaches rely strongly on the quality and quantity of training data. In real-world applications, however, it is challenging to obtain sufficient data suitable for these approaches. To alleviate this problem, a novel deep generative model for SAR image generation is proposed, which is an extension of Wasserstein autoencoder. The network structure and reconstruction loss function of the model have been improved according to the characteristics of SAR images. The experimental results demonstrate that our model is superior to other classical generative models in SAR image generation. The generated images can be directly used as training samples, thereby extending the training data set and improving the recognition accuracy.
10 April 2020 in IEEE Geoscience and Remote Sensing Letters
Numerous synthetic aperture radar-automatic target recognition (SAR-ATR) methods require a large amount of training data. However, collecting SAR data is both expensive and complicated in practical applications. Recognition with the limited training data has become a vital issue in SAR-ATR. To solve this problem, we propose a recognition model combining probabilistic inference with meta-learning to transfer prior knowledge from simulated to real SAR data. First, we use various recognition tasks drawn from the simulated data to learn the global parameters of the model. Second, we draw new tasks from the real data and use the amortized inference to model a posterior distribution over task-specific parameters. Finally, we produce a predictive distribution indicating the confidence of the target classes. The experimental results demonstrate the superiority of the model in recognition tasks with a small amount of training data. We also show that introducing probabilistic inference can improve the prediction accuracy and prediction uncertainty of the model.
17 October 2025 in IEEE Access
Few-shot image classification (FSIC) is a critical task in computer vision that aims to accurately classify new categories with only a limited number of labeled examples. This capability is essential for practical applications such as medical diagnosis, remote sensing, and autonomous driving, where data scarcity is common. Although numerous FSIC methods have been proposed, a comprehensive systematic literature review (SLR) remains absent. This review begins by summarizing foundational FSIC approaches and evaluating their effectiveness across standard benchmarks. It then explores recent advancements from 2019 to 2025, providing an in-depth analysis of key factors affecting FSIC performance, including dataset characteristics, model architecture, and evaluation protocols. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, multiple academic databases were searched, resulting in the selection of 26 relevant studies based on strict inclusion and exclusion criteria. The findings highlight that FSIC performance is strongly influenced by factors such as the number of support samples, domain shifts, label noise, and model complexity. Finally, the review identifies future research directions, emphasizing the integration of domain adaptation, lightweight model design, and multimodal learning, while addressing current challenges to guide the next phase of FSIC development.
15 July 2025 in IEEE Access
Deep learning, with large-scale annotated datasets, has demonstrated remarkable success in synthetic aperture radar automatic target recognition (SAR-ATR). However, the collecting of SAR images is expensive and complex, and manually labeling them requires expert knowledge. To overcome these limitations, we propose a few-shot learning model capable of accurate recognition of novel targets with minimal training samples. Our model innovatively integrates causal inference with mutual centralized learning (MCL) to address few-shot SAR-ATR tasks. First, we establish a causal inference framework to identify and model the dependencies among target characteristics, imaging conditions, and category labels. This framework incorporates a novel causal intervention method based on multi-scale random convolution to eliminate spurious correlations caused by imaging variations, thereby enhancing feature stability. Second, we introduce an advanced MCL module to effectively evaluate feature similarity in few-shot settings. MCL breaks through the unidirectional matching paradigm adopted by conventional metric learning. Through its bidirectional feature interactions and dense feature accessibility mechanisms, MCL achieves more robust feature discrimination in few-shot learning tasks. Comprehensive experiments demonstrate that our model outperforms existing advanced few-shot SAR-ATR methods, achieving superior recognition accuracy while maintaining robustness in data-scarce scenarios.