In response to the limited detection ability and low model generalization ability of the YOLOv7 algorithm for small targets, this paper proposes a detection algorithm based on the improved YOLOv7 algorithm for steel surface defect detection. First, the Transformer-InceptionDWConvolution (TI) module is designed, which combines the Transformer module and InceptionDWConvolution to increase the network's ability to detect small objects. Second, the spatial pyramid pooling fast cross-stage partial channel (SPPFCSPC) structure is introduced to enhance the network training performance. Third, a global attention mechanism (GAM) attention mechanism is designed to optimize the network structure, weaken the irrelevant information in the defect image, and increase the algorithm's ability to detect small defects. Meanwhile, the Mish function is used as the activation function of the feature extraction network to improve the model's generalization ability and feature extraction ability. Finally, a minimum partial distance intersection over union (MPDIoU) loss function is designed to locate the loss and solve the mismatch problem between the complete intersection over union (CIoU) prediction box and the real box directions. The experimental results show that on the Northeastern University Defect Detection (NEU-DET) dataset, the improved YOLOv7 network model improves the mean Average precision (mAP) performance by 6% when compared to the original algorithm, while on the VOC2012 dataset, the mAP performance improves by 2.6%. These results indicate that the proposed algorithm can effectively improve the small defect detection performance on steel surface defects.
Read full abstract