Joint attention mechanism with dynamic kernel for yolov5 mobile wireless charging coil surface defect identification.

The yolov5-CTD (you only look once version five-carafe triplet double attention vision transformer) coil defect detection algorithm is proposed to address the problems associated with the manufacturing process of wireless charging coils, which are the core component of wireless chargers and can produce multiple types of defects. To address the problem of small area defects, a large amount of detail semantic information is lost in the process of down-sampling in the backbone network. DA-ViT (Double Attention Vision transformer) is incorporated into the enhanced feature extraction network to supplement the detail semantic information and enhance the ability of the network to build up the extraction of long-distance information. The Triplet attention mechanism module is introduced to be embedded in the lateral hop connection of the multiscale feature extraction network to enhance the neck network's ability for local information extraction by rotating the dimension for multidimensional feature capture and improve the effect of multiscale feature fusion. For the large number of extreme aspect ratio defects in the coil, Carafe up sampling is used to aggregate contextual information to improve the perceptual field, while optimising the jaggedness and mosaic phenomenon of the defect edges of the feature map caused by nearest neighbor interpolation up sampling. To enable the backbone feature extraction network to learn better feature information, the convolution part of yolov5 is optimised to dynamic conditional convolution. Experiments have shown that yolov5-CTD can detect surface defects on mobile phone wireless charging coils with the accuracy of Map@0.5 reaches 80.9% which is 4.8% higher than the original network, and the detection speed is 41.15FPS, which can meet the industrial production line requirements in terms of speed and accuracy. [ABSTRACT FROM AUTHOR]