Research on scratch defect detection method based on attention mechanism

Scratches are subtle surface defects with various sizes and shapes. Texture noise in industrial environments can easily cover up or confuse the defects. In addition, uneven lighting, shadows, and dirt existing in the actual production environment will also affect the visual characteristics of scratches. Therefore, this article proposes a lightweight semantic segmentation network based on the deepLabv3+ model. First, in order to reduce the computational complexity of the model and improve computational efficiency, we used the more lightweight MobileNetV2 network to replace the backbone feature extraction network Xception in the original model. Secondly, combined with the slender characteristics of the scratch itself, this paper proposes the SP_ASPP module, which adds Strip pooling to the ASPP structure for more accurate feature extraction. This article selects the lead frame data set for verification. The experiment shows that the improved model surpasses the original model in terms of speed and accuracy, and the miou value of the lead frame reaches 85.86%.