Detection of small defects on gear surface based on improved YOLOv7

In industrial quality inspection, false inspection or undetected inspection often occur in the identification of gear surface minute defects. This paper presents a novel detection framework for identifying surface defects on gears, utilizing an enhanced version of YOLOv7. Firstly, we add a small target detection layer upon the foundation of YOLOv7 to enhance the network’s capability of accurately detecting smaller targets. Secondly, we replace the MP2 module with the MPCA module that integrates the CA (Coordinate Attention) mechanism, which can reduce the amount of computation and increase the network’s attention to small targets by suppressing the interference of unnecessary information. Finally, for small and dense defects on the gear surface, the original Binary Cross Entropy Loss and CIoU Loss are substituted with the Varifocal Loss and SIoU Loss to train the network. The empirical findings demonstrate that the improved model significantly enhances the precision and recall rate of small defects detection. mAP@0.5 achieves 77.7%, exhibiting a notable improvement of 3.3% over the baseline YOLOv7 model, and better than other one-stage target detection networks.