GAN-based industrial part optical image generation model with controllable constraints

The generation of industrial part optical image is particularly important for industrial vision inspection. However, the current digital measurement technology represented by visual measurement has many problems: a) the lighting debugging work requires a large amount of experience from visual engineers, b) the system development cycle is long and does not meet the flexibility requirements, c) the adaptability is poor, and it is difficult to adjust different parts quick response to differentiated testing. Therefore, this paper proposes an image generation model with controllable constraints added based on Generative Adversarial Networks, which can generate information such as illumination distribution and texture of the industrial part optical image. We use an image conversion model based on Generative Adversarial Networks, and add label vectors as controllable constraints to generate industrial part optical images with different texture and illumination distribution. In addition, a gradient penalty term is introduced in the loss function to stabilize model training. Experiments show that the gray distribution of the generated image fits the gray distribution of the real image better, and can reflect the real texture and brightness trend of the image. Compared with previous image acquisition and image rendering methods, our method can quickly obtain industrial part optical image under different conditions without tedious lighting debugging.