A Visual Inspection and Classification Method for Camshaft Surface Defects Based on Defect Similarity Measurement

Industrial parts are usually shaped with complex surface structures and they require high surface quality. The surface defects influence the precision and performance of them. The efficient optical inspection equipment and defect detection and classification method are important for improving t productivity. There are two main difficulties in the visual detection of surface defects of industrial parts. One is difficult to develop visual inspection equipment for complex surfaces. Second is that the optimal feature set extracted by the commonly used feature fusion classification methods is not sensitive to defects with high defect similarity, which affects the defect classification accuracy. To solve difficulty 1, We have researched the principles of refraction and reflection when a light beam reaches the interface of a non-homogeneous medium, and develop a high-performance optical inspection equipment, which is available for 360° image acquisition of the external surface of an industrial part. To solve difficulty 2, a feature fusion classification method based on defect similarity measurement is proposed. The main idea is to measure the similarity between various types of defects by calculating the Euclidean distance between class centers and to classify defects with similarity higher than a threshold into one category. Subsequently, a two-step feature extraction and classification strategy is adopted. In the first step, the lower similarity types are separated, and then in the second step, the higher similarity types are separated. Deep learning methods and traditional machine learning methods are used for validation, respectively. Experimental results indicate that the proposed method achieves a 5% improvement in overall classification accuracy, and also has significant improvements in precision, recall and many other indicators compared to several other advanced classification methods.