DFP-YOLO: a lightweight machine tool workpiece defect detection algorithm based on computer vision

Achieving accurate identification of various defects in workpiece surface defect detection is a complex task. To address the issues of high computational complexity and model complexity in existing defect detection algorithms, this paper proposes an improved network model algorithm, DFP-YOLO, for defect detection on cutting tool workpieces. The improved algorithm first introduces the C3_D module and C3_F module to reduce model parameters and complexity. Secondly, the bounding box regression loss function utilizes SIoU to improve detection accuracy. To address the issues of imbalanced detection samples and long training time, transfer learning is employed to accelerate the convergence speed of the overall network model and save computational resources. Finally, channel pruning algorithm is applied to compress the overall network structure without significantly affecting detection accuracy, resulting in a more lightweight model. The experimental results demonstrate that DFP-YOLO effectively reduces the parameter quantity by 51% and GFLOPs by 50% while preserving the mean average precision (mAP) unchanged. Moreover, the detection speed remains constant at 108 frames per second. This improved algorithm surpasses existing methods for surface defect detection and achieves both fast and accurate detection of defects on workpieces. Code is available at https://github.com/wdikky/DFP-YOLO