SDPS-M²CAN: Improving predictive performance of Photometric Stereo Networks on broad spectrum reflective surfaces with minor indentations or bumps

Photometric stereo is a method used to estimate surface normals of objects, commonly employed in detecting surface defects on industrial products. To enhance the detection of surface features on smoother and more continuous industrial components, this study improves the learning-based photometric stereo approach, SDPS-Net, resulting in the model SDPS-M²CAN. Firstly, we introduce the Multi-scale Channel Attention Module (MCAB) to optimize the extraction of photometric stereo features, allowing the model to effectively adaptively extract useful features across multiple scales. Secondly, in the upsampling network, we employ a multi-scale feature fusion strategy based on maxpooling to fully utilize photometric stereo features at different scales, thereby enhancing the prediction capability of surface normals. Finally, by incorporating the ResNet architecture to simplify the model training process and expanding the network channels, we further improve prediction accuracy. Quantitative and qualitative experiments demonstrate that our enhanced model outperforms the original SDPS-Net in predicting concave-convex areas and normals near edges of smooth and continuous broad-spectrum reflective materials, achieving more accurate overall normal predictions.