Research on tire appearance defect detection algorithm based on efficient multi-scale convolution

Abstract Due to the large randomness of tire appearance defect size and the complex and diverse defect shapes, the existing target detection algorithm is prone to missing and misidentifying targets, the accuracy is limited, and the detection model is large, which is not conducive to deployment on embedded devices. In this paper, the efficient multi-scale convolution (EMC) mode is proposed, and the C2f-EMC module is designed on this basis, which improves the network structure of YOLOv8, improves the accuracy of tire appearance defect detection, and reduces the number of parameters in the model. EMC convolution first divides the input feature images into four parts on average and carries out multi-scale convolution with convolution cores of 1×1, 3×3, 5×5 and 7×7 sizes respectively. Then, the obtained results are stacked, and cross-channel feature fusion is realized by point-by-point convolution. After determining the network structure of C2f-EMC, the best improvement position of C2f-EMC module is determined through comparative experiments. Experiments show that after the above improvements, the parameter number of the model is reduced by 4.85%, the calculation amount by 2.82%, the model size by 4.44%, the recall rate by 2.8%, the mAP50 by 1.0%, the mAP50-95 by 1.3%, and the F1 by 2%. The defect detection task can be completed more accurately and the model size requirements of embedded devices can be better met.