A Conjugate Gradient Neural Network for Inspection of Glass Defects

To solve the problem in projecting grating method, this paper presents an inspection method combining phase map and characteristic image for glass defects. By using this method, the phase difference is achieved according to the fringe images of defect-free and defect-containing, meanwhile the characteristic image of defects-containing is also obtained by 1D Fourier transform. The segmentation of defect region is implemented by integrating grayscale mathematical morphology with threshold segmentation, and the boundary coordinate of connected region is used to calculate the size and location of defect. The defect region in characteristic image is extracted correspondingly according to the boundary coordinate of connected region. The second iteration segmentation method based on grey range is applied to calculate the low and high thresholds, and acquired the ternary-valued defect image. A Conjugate Gradient Neural Network (CGNN) is designed to recognize the type of defect, and the accuracy of the recognition reaches 86%. The results of typical defects demonstrate that the proposed method provides reliable identification of defects.