Defect Sample Image Generation Method Based on GANs in Diamond Tool Defect Detection

The method of diamond tool defect detection using deep learning technology is a novel algorithm in engineering applications. The method combining machine vision and image processing can realize the intuitive and rapid detection of tool defects. Obtaining good performance using these methods largely relies on a considerable number of training samples, which hinders their wide application in the field of diamond tool defect detection. This paper presents an approach designed using small sample data to address this issue. First, a data augmentation algorithm is implemented by generative adversarial networks (GANs) based on existing defective data samples and the principle of image-to-image translation. With the support of the trained network model, numerous defect-free diamond tool images are transformed into diamond tool images with defects. Besides, the image generation results in three annotation modes are compared, and the image generation results that perform well are selected. According to the characteristics of image fusion, some generated defects are combined with the original image to synthesize the fusion-generated image with a better effect. Then, a large data set composed of generated samples combined with real tool defect image samples is employed to detect diamond tool defects. The method is applied to the experiment of diamond tool sharpening quality detection, and the average accuracy of network recognition can be improved by 8.4% after using the augmented data set. Compared with the data augmentation method using image cutting and flipping, the method proposed in this paper also has a 1.6% improvement.