LAD-YOLO: A Lightweight YOLOv5 Network for Surface Defect Detection on Aluminum Profiles

In this paper, we leverage the advantages of YOLOv5 in target detection to propose a highly accurate and lightweight network, called LAD-YOLO, for surface defect detection on aluminum profiles. The LAD-YOLO addresses the issues of computational complexity, low precision, and a large number of model parameters encountered in YOLOv5 when applied to aluminum profiles defect detection. LAD-YOLO reduces the model parameters and computation while also decreasing the model size by utilizing the ShuffleNetV2 module and depthwise separable convolution in the backbone and neck networks, respectively. Meanwhile, a lightweight structure called "Ghost_SPPFCSPC_group", which combines Cross Stage Partial Network Connection Operation, Ghost Convolution, Group Convolution and Spatial Pyramid Pooling-Fast structure, is designed. This structure is incorporated into the backbone along with the Convolutional Block Attention Module (CBAM) to achieve lightweight. Simultaneously, it enhances the model's ability to extract features of weak and small targets and improves its capability to learn information at different scales. The experimental results show that the mean Average Precision (mAP) of LAD-YOLO on aluminum profiles defect datasets reaches 96.9%, model size is 6.64MB, and Giga Floating Point Operations (GFLOPs) is 5.5. Compared with YOLOv5, YOLOV5s-MobileNetv3, and other networks, LAD-YOLO proposed in this paper has higher accuracy, fewer parameters, and lower floating-point computation.

@article{Zhao2023,
  title     = {LAD-YOLO: A Lightweight YOLOv5 Network for Surface Defect Detection on Aluminum Profiles},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {9},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Dongxue Zhao and Shenbo Liu and Yuanhang Chen and Da Chen and Zhelun Hu and Lijun Tang},
  doi       = {10.14569/IJACSA.2023.0140924},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140924}
}