HVA-YOLO: A real-time instance segmentation method under Hybrid Visual Attention applied to steel plate surface defect detection

End-to-end real-time industrial surface defect detection is a crucial aspect of the production process for mechanical systems. Numerous studies have employed computer vision methods in defect signal processing tasks. Defects on the surface of industrial products often have large feature scale spans, diverse features and randomization, resulting in a single horizontal target detection frame that does not represent the defects well. The instance segmentation method can represent the defect region pixel by pixel, which is also helpful for subsequent monitoring and process control. This study further optimizes the feature condition self-recognition module based on the YOLO network. Multi-head self-attention is utilized to constitute an efficient visual self-attention mechanism to capture further the contours and detailed information of defects for feature learning with hybrid visual attention. Hybrid pooling and depthwise separable convolution for spatial channel down-sampling module are proposed for the first time and applied to the backbone and neck stages. The module can further enhance the ability of neural networks to extract high-level semantic information and effectively reduce computational resources. The experimental results show that the proposed method can achieve significant detection accuracy and speed improvement in steel plate surface defect recognition. Meanwhile, the real-time instance segmentation method not only empowers intelligent manufacturing, but also the segmented defect regions can help the subsequent improvement of the production process to realize the intelligence of the mechanical manufacturing system.