Attention mechanism and texture contextual information for steel plate defects detection

In order to achieve rapid inference and generalization results, the majority of Convolutional Neural Network (CNN) based semantic segmentation models strive to mine high-level features that contain rich contextual semantic information. However, at steel plate defects detection scenario, some background textures’ noises are similar to the foreground leading to hard distinguishment, which will significantly interfere with feature extraction. Texture features themselves often hold the most plentiful contextual information. Despite this, semantic segmentation tasks rarely take texture features into account when identifying surface defects on steel plates. In that case, the essential details, such as the edge texture and other intuitive low-level features, will generally cannot be included into the final feature map. To address the problems of inefficient accuracy and slow speed of existing detection, this study proposed a steel plate surface defect detection method using contextual information and attention mechanism, and utilizes a multi-layer feature extraction method and fusion framework based on low-level statistical textures. Through the identification of pixel-level spatial and correlation relationships, characteristics of low-level defects are extracted. Furthermore, to effectively incorporate statistical texture in CNN, a novel quantization technique has been developed. This quantization method allows for the conversion of continuous texture into various levels of intensity. The network parameters were iterated in a gradient direction, facilitating the defects division. Empirical results have demonstrated the feasibility of applying the proposed approach to practical steel plate testing. Additionally, ablation experiments have demonstrated that the method is capable of effectively enhancing surface defect detection for steel plates, resulting in industry-leading performance.