GSD-YOLO: A Gear Surface Defects Detection Method Using Adaptive Multiscale Fusion and Hybrid Feature Fusion

Gear surface is the main working interface of the gear, which directly influences the safety and lifespan of mechanical equipment with different gear surface defects. Due to the large-scale variations, diversified types, multiple faults overlapping, noise interference and low contrast between background and defects, the GSD detection is prone to occur false detection or missed detection. To address the issues and precisely identify the gear surface defect, a novel GSD-YOLO network based on YOLOv5 is proposed. Firstly, an AMF module is constructed and used to displace the C3 module in the neck structure, which can enhance the feature extraction capability of the neck for multi-scale and multi-type GSD. Then, the DFE module is integrated into HFF module. Afterward, the HFF module is added before the detect layer and the recognition ability for low contrast and overlapping defects is improved. Finally, the random noise is introduced in the data augmentation process to reinforce anti-noise performance. Experimental analysis shows improvements in mAP by 2.3% and 3.0% on NEU-GSD and RSDDs datasets, respectively. Meanwhile, the improved GSD-YOLO shows stronger robustness and generalization ability when dealing with complex defects, and the mAP values reach 96.3% and 82.1% on the two datasets, which are better than other advanced models.