Intelligent detection of surface defects in sheet metal parts based on photometric stereo vision

Surface defect detection of sheet metal parts is a crucial component in achieving closed-loop control of stamping production quality and ensuring overall product reliability. Machine vision-based quality inspection techniques have been increasingly adopted for surface quality evaluation across various products. However, Surface defect detection of sheet metal parts is often constrained by challenges such as complex surface geometries, intense specular reflections, stringent defect detection standards, and limited model performance, which collectively hinder effective surface defect inspection. To address the challenges in surface defect detection of sheet metal parts, this paper proposes an detection method based on photometric stereo vision and deep learning. By utilizing multi-light source illumination and photometric stereo algorithm, normal vector images that go beyond traditional 2D grayscale images are obtained, enabling a multi-dimensional representation of surface defects. To enhance feature representation and detection performance, the ODConv convolutional module and CBAM attention mechanism are integrated into the YOLOv8 model. Based on these improvements, a defect detection framework utilizing normal vector images was developed. Experimental results demonstrate that this method achieves a detection accuracy of over 93.3% for defects such as surface unevenness in automotive sheet metal, significantly outperforming the original model. This approach overcomes the limitations of conventional 2D detection and offers a promising solution for high-precision surface defect inspection of sheet metal parts.