Intelligent Detection System of Mechanical Parts Surface Defects based on Computer Vision

This study addresses accuracy and real-time challenges in mechanical part surface defect detection by proposing an improved YOLOv8 system. Integrating CBAM (Convolutional Block Attention Module) attention mechanisms enhances small-target feature extraction, while TensorRT FP16 (Half-Precision Floating Point) quantization enables 41.2 FPS inference on NVIDIA Jetson devices (46.3% faster than FP32, 0.3% accuracy loss). A dynamic learning rate strategy and OHEM (Online Hard Example Mining) improve complex scene adaptability. Evaluated on NEU-DET (Northeastern University Defect Detection Dataset), the system achieves 83.2% mAP, 89.2% small-target recall, and <1.8% false positives, outperforming YOLOv5s and Faster R-CNN in precision-speed-robustness balance. Future work will focus on unsupervised domain adaptation for extreme lighting and irregular defects.