Abstract: To address the issues of low efficiency and accuracy in detecting numerous complex micro defects on the surface of parts, an improved detection network called BCGS-YOLOv7 tiny is proposed, based on the YOLOv7-tiny architecture. In the feature extraction stage, the CBAM (Convolutional Block Attention Module) attention mechanism is incorporated, and the SPD-Conv downsampling layer is used instead of stride convolution and max-pooling layers to enhance the feature extraction capability for small objects. In the feature fusion stage, the Path Aggregation Feature Pyramid Network (PAFPN) is replaced by the BPANet network structure, and the gnCSP module and SPD-Conv downsampling layer are introduced to improve the feature fusion ability for small objects.Experimental results on the reconstructed GC10-DET dataset demonstrate that the BCGS-YOLOv7 tiny detection network achieves a mean Average Precision (mAP) of 91.6%, which is a 6.0% improvement over the original YOLOv7-tiny network. Moreover, the detection accuracy for various types of micro defects on part surfaces is significantly enhanced.