A novel approach for surface defect classification based on 3D point cloud data with hybrid filters

Surface defect identification is one of the critical topics in manufacturing for ensuring product quality and performance. The conventional methodology of surface defect identification applies the visible light imaging technique, but struggles with variations in lighting conditions. The lack of depth information of the visible light imaging technique makes the detailed surface morphology analysis and classification task challenging. This study proposes a novel method leveraging 3D point cloud data and corresponding surface normal vectors to classify surface defects. A hybrid filtering approach, combining Gaussian and Gabor filters, is applied to the surface normal vectors to enhance defect-related features extraction while mitigating noise. The processed vectors are then transformed into RGB images, enabling the use of ResNet 18 for effective feature extraction and classification via transfer learning. Experimental results demonstrate that the proposed method achieves 90% accuracy in classifying surface defects with real-world datasets. Compared to existing approaches, the proposed method eliminates the dependence on lighting conditions while leveraging the advancement of Convolution Neural Network (CNN) in the field of image processing. This research contributes a scalable and efficient framework for surface defect inspection in modern manufacturing environments.