Defect Detection on the Surface of the Bellow Expansion Joints Based on Machine Vision

Since the surface quality of the expansion joints has a large impact on bellows performance, the detection of surface defects in expansion joints is an important aspect of their production process. This paper presents a two-step method for detecting unpredictable faults based on machine vision. The expansion joints are first segmented from the smooth parts of the bellows and background using a single Gabor filter. The second step is to split the segmented expansion joint into blocks using the Haar feature response, and then the features in each block extracted with the Haar feature are normalized to represent the surface condition of the block. A number of defect-free samples were used to obtain the normal range of normalized features. Experiments were performed on 10 mm diameter plastic bellows and compared with deep learning methods GAN and CFLOW-AD. The accuracy of our proposed method was verified according to experimental tests using 500 images from real samples. Results show that the proposed method can realize real-time detection with high accuracy in the industry.