Yolov7-pcbam: enhancing steel surface defect detection via partial convolution and attention mechanism

Abstract Surface defect detection on steel plates remains a critical challenge, primarily due to the inefficiency and subjectivity associated with manual inspection. To overcome this limitation, an enhanced object detection model is proposed based on YOLOv7, a state-of-the-art real-time object detection framework. The improved model, named YOLOv7-PCBAM, incorporates a series of architectural modifications to enhance detection performance. We modify the efficient layer aggregation network (ELAN) and its associated blocks, which are core components in YOLOv7, by replacing most of their standard convolutions with partial convolutions (PConv). This novel modification reduces model complexity while preserving effective feature aggregation. To further boost detection accuracy, the convolutional block attention module (CBAM) is incorporated into the neck of the network, making feature extraction more comprehensive and effective. The Mish activation function is also employed in the convolution–batch normalization–SiLU (CBS) blocks to replace the original SiLU, which contributes to improved feature representation and stable training by mitigating gradient vanishing. The proposed model is evaluated on two widely used steel defect datasets, NEU-DET and GC10-DET, and is compared with the original YOLOv7 and YOLOv9 as well as the lightweight variants YOLOv7-tiny and YOLOv8-n. The results demonstrate that YOLOv7-PCBAM consistently improves most evaluation metrics over YOLOv7 and YOLOv9 on both datasets and attains higher [email protected] and recall than YOLOv7-tiny and YOLOv8-n. In particular, it achieves 4% and 3% improvements in [email protected] over YOLOv7 on NEU-DET and GC10-DET, respectively, while reducing computational complexity in GFLOPs by 43% and maintaining high detection accuracy. The high recall of YOLOv7-PCBAM indicates a lower risk of missed defects, which is essential for reliable industrial inspection.