Abstract:

Aiming at the problems of insufficient feature extraction ability, loss of detail information, and low computational efficiency in the detection of minor damages and unobvious damages on the car body paint surface by Yolov7-Tiny(You Only Look Once v7 Tiny), an improved Yolov7-Tiny algorithm is proposed. Firstly, the RepNCSPELAN4 ( Re-parameterizable Cross Stage Partial Efficient Layer Aggregation Networks) module is introduced into the backbone network to replace the ELAN-Tiny(Eficienta Layer Aggregation Networks Tiny)module, to enhance the feature extraction ability of the model. Secondly, the Swish activation function is introduced to improve the nonlinear expression ability of the model. Finally, the SPDConv ( Space to Depth Convolution) is introduced into the neck network. Combined with the improved SPDConv_Res(Space to Depth Convolution _ Residual ) detection head, it retains the detail information of small targets and balances the parameter burden. The DSConv(Distribution Shift Convolution) is introduced into the neck network, ensuring the full fusion of defect feature information and improving the computational efficiency simultaneously.Experimental results on the self-built dataset show that the mAP@ 50 of the improved model reaches 86. 0% ,which is 7. 4% higher than that of the baseline model, and the computational amount is reduced by 3. 3 x10⁹ times. While improving the accuracy, it also reduces the demand for computational resources to a certain extent.

Key words: