Solid Surfaces, Defects and Detection

In the modern industrial field, particularly in steel and automobile manufacturing, detecting defects in steel surfaces is crucial to product quality and safety. Due to the complexity of the manufacturing process and the variety of defect types, traditional detection methods have struggled to meet the demand for efficient, accurate, and real-time detection. In recent years, the rapid development of deep learning, computer vision, and image processing technologies has provided new solutions for steel surface defect detection. These new technologies not only enhance detection accuracy but also significantly accelerate detection speed, promoting the advancement of industrial automation.

In their paper, “Steel Surface Defect Recognition: A Survey”, Wen et al. conduct a comprehensive review of the development history of steel surface defect recognition technology and systematically summarize the progression from traditional image processing to deep learning methods [1]. They compare the differences between traditional and deep learning methods in terms of detection accuracy, real-time performance, computational complexity, etc. and summarize the main challenges in the field of steel surface defects detection, which are insufficient and unbalanced data samples, real-time detection, and small object detection. Finally, they provide appropriate recommendations for resolving these issues. By summarizing and comparing a large amount of literature, their paper not only provides readers with a comprehensive view of the field, but also provides theoretical support and guidance for subsequent research. The paper has been selected as an ESI highly cited paper on multiple occasions.

In practical inspection scenarios, due to the blurriness and low resolution of steel defect images, the features learned by the network suffer from information loss, feature blurring, and confusion, which makes it difficult to recognize small defects. Guo et al. proposed using ASENet, an autocorrelation semantic enhancement network, for the classification of steel surface defects. This network extracts the basic features of an image through a backbone network, enhances them using a CS attention module, and calculates the correlation between neighboring features using an autocorrelation module. Finally, the enhanced features are combined with the basic features via residual linking to obtain self-attention features, which enhance the semantic information of the image and improve defect recognition accuracy for low-resolution images [2]. Similarly, X-ray inspection, an important nondestructive testing method, is utilized for the detection of tiny defects. Such detection relies on high-resolution imaging techniques. Liu et al. improved the resolution of X-ray imaging using sub-pixel displacement, which can clearly recognize tiny defects [3]. Combining this method with deep learning techniques facilitates more accurate steel surface defect detection. Liu et al. also designed the UNet++ network with a variable attention mechanism and incorporated a preprocessing method based on the pyramid model, which further improves the performance and visibility of the extraction of faint defects, and provides effective high dynamic range compression and defect enhancement of the 16-bit raw image, improving the detection accuracy of steel surface defects in the field of nondestructive testing [4].

Traditional methods for classifying steel surface defects improve accuracy by increasing the depth of the convolutional neural network (CNN) and the number of parameters, but this approach ignores the memory overhead of large models and the incremental gain in accuracy decreases as the number of parameters increases. To address this problem, Shao et al. proposed a multi-scale lightweight neural network (MM), which takes the fusion coding module as the core and uses a Gaussian difference pyramid to construct a multi-scale neural network, which realizes the effective extraction of features at different scales and adapts to a variety of defect types [5]. In resource-constrained industrial scenarios, this method ensures detection is accurate and efficient. Similarly, to improve the detection accuracy and efficiency, Huang et al. proposed WFRE-YOLOv8s for steel surface defect detection based on YOLOv8s [6]. They addressed the data quality imbalance and designed a new neck module, RFN, which reduces computational overhead and effectively fuses features from different scales. Finally, through the EMA attention module, valuable features are more effectively extracted, improving both the detection accuracy and speed of the model and optimizing steel surface defect detection performance.

During steel surface defect detection, the dataset often suffers from a severe class imbalance, which degrades the performance of traditional models. Yu et al. proposed a distribution-preserving undersampling method, which divides all normal samples into several subgroups through cluster analysis and recombines them into balanced datasets. This ensures that the normal samples in all balanced datasets have the same distributions as the original imbalanced dataset, effectively resolving the data imbalance [7]. On the other hand, He et al. enhanced model training using semi-supervised learning by leveraging a large amount of unlabeled data in the case of label scarcity. This approach improves defect detection accuracy for complex surface textures. Although the two papers address different data challenges, both enhance the generalizability and robustness of the model through innovative learning strategies [8].

In practice, due to the randomness of defects, their textures and shapes are often very similar to the background, making defects difficult to recognize. To address this issue, Wan et al. adopted a dual-stream Swin Transformer network architecture to process local and global features separately, further improving detection accuracy [9].

In addition to steel surface defect detection, some research has extended defect detection technology to other industrial fields. Liu et al. proposed a defect detection method applicable to automobile body-in-white, which can handle the detection of complex surface features under different working conditions [10]. This study demonstrates the potential for cross-industry application of defect detection technology, providing a reference for other industrial fields.