GAN-based Method with Small Training Samples for Defect Detection Using Magneto-optical Image

Due to the large amount of noise in the magneto-optical image, it is difficult to select the appropriate threshold to segment the image. To solve this problem, the characteristics of the pixel distribution of magneto-optical image are taken into consideration and a small sample training method based on generative adversarial network is proposed to enhance the contrast of the magneto-optical image. First, an excitation coil is used to excite the specimen with defect, and the magneto-optical image of the specimen is obtained by magneto-optical imaging instrument. Second, after normalized, the original magneto-optical image is sampled randomly to generate sufficient training data. Then, the generated data are used to train the generative adversarial network. Finally, the testing image is sent to the trained network for enhancement and then segmented. The experimental results show that the data expansion method proposed in this paper is suitable for small sample training, and the network after training effectively improves the contrast of magneto-optical image, making it easier to select the segmentation threshold.