Abstract: In the domain of automotive paint defect detection, to address the issue of defect feature information loss due to specular reflection on the surface, bimodal data samples were created using phase measurement deflection techniques to enhance defect features and reduce background complexity, focusing on the location of foreground defects. Additionally, to tackle the low efficiency of manual detection methods and the low accuracy of traditional detection methods, improvements were made to YOLOv8. Self-correcting convolutions and an ADown downsampling module were introduced in the backbone network to enrich and enhance the information flow. A receptive field attention convolution replaced the ordinary convolution in the original neck to focus on important features. The original neck's upsampling was replaced with the DySample module to reduce feature loss during spatial scale transformation. Experimental results reveal that the proposed method excels in detecting three specific types of automotive paint defects: dents, bumps, and scratches. Compared to the YOLOv8n baseline model, it achieves a 1.9% improvement in both mAP50 and mAP50-90 metrics. Furthermore, the improved YOLOv8 method has been validated on the NEU-DET and PASCAL VOC 2012 public datasets, demonstrating strong applicability for detecting objects with various attributes.