Application of attention YOLOV 4 algorithm in metal defect detection

Common feature engineering method and traditional machine visual detection algorithm have problems with strong subjective dependence, low detection accuracy and limited detection range in the detection of metal surface defects. Integrated the ECA attention mechanism to realize the adaptive weight assignment in the important areas of the image will form ECAMobileNetV2 as the model backbone feature extraction network, then use the PANet module of YOLOV4 to enhance the defect feature-one lightweight Yolo V 4 model (ECA_MobileNetV2_yoloV4, abb EMV2yoloV4) integrated ECA and MobileNet. Our method got highest detection accuracy, applied the datasets of metal surface defects for defect types in GCT10 and NED_DET, with mAP of 0.86 and 0.68 respectively. it's significantly higher than MV2yoloV4 and MV3yoloV 4 integrating attention mechanism SE. The model parameter reaching 10.4M is less lightweight than novel detection networks such as Efficientdet and Ghost etc. Experexperiment shows that EMV2yolo V 4 better solves the problem of low recognition accuracy caused by background pixels and brightness. The single image inference time of 18.44ms and frame rate up to 54.25f/s. It can meet the requirements of lightweight deployment and accuracy requirements of metal surface defect detection.