Steel Surface Defect Detection Method Based on SEE-YOLO

This study proposes an efficient steel surface defect detection algorithm, i.e., SEE-YOLO, to overcome the problems of insufficient multiscale feature extraction, low detection accuracy, and high model complexity. First, a crossstage partial (CSP) bottleneck with two convolutions incorporating a Sobel-conv fusion module (C2f_SCFM) was added to the backbone network to fuse edge and convolution features, leading to a notable enhancement in the model's feature extraction performance. Second, an efficient upsampling and channel block (EUCB) was introduced to the neck network to optimize feature upsampling efficiency. Finally, an efficient multibranch and scale feature pyramid network (EMBSFPN) is proposed, which effectively preserves shallow-layer information, enriches the gradient information of the output layer, expands the receptive field using reparameterized large convolutions, and dynamically extends the network's effective receptive field, while retaining information on small targets.In addition, In the BiFPN architecture, an adaptive weighted fusion mechanism was employed to lower model complexity while preserving performance accuracy. Experimental results on the NEUDET dataset showed that precision (P) increased by 4.3 %; with a 3.8% gain in mAP50 and corresponding reductions of 37.8 % and 16 % in model parameters and computational requirements, respectively, the method achieves a superior balance of precision and efficiency. The proposed model performed well on the PKU-PCB dataset, yielded comprehensive performance in steel surface defect detection, and outperformed some state-of-the-art detectors.