Improved YOLOv7-tiny algorithm for surface defect detection of bushing parts

For improving the precision and efficiency of sleeve surfacedefect detection, an improved YOLOv7-tiny algorithm is proposed in this paper. First, in the feature extraction of the model, the standard convolution is replaced by omni dimensional dynamic convolution for processing data of any dimension, and second, in the feature fusion, the standard convolution is replaced by omni dimensional dynamic convolution, the nearest neighbor interpolation of the upper sampling part is replaced by the light-weight operator CARAFE, the BiFormer is introduced into the splice to increase the detection of local small targets, and the Slim-Neck module is introduced by replacing the standard convolution with GSConv. Finally, compared with the original model, the modified algorithm improved 7.7% on mAP, 11% on local small targets and 40.3 on FPS. Experiments on GC10-DET data set show that the improved algorithm is universal.