A denoising and restoration method of weld laser stripe image for robotic multi-layer multi-pass welding based on generative adversarial networks

Abstract

Obtaining clean, intact, and clear laser stripe images from weld laser stripe images with noise is a key issue in robot welding systems based on laser vision sensors, especially for robotic multi-layer multi-pass welding. However, existing methods typically require image pairs, consisting of images with noise and corresponding images without noise, which are very difficult to obtain. Meanwhile, methods that do not require image pairs suffer from low accuracy. Therefore, a model called DARM (denoising and restoration model) based on CycleGAN is proposed to denoise and restore laser stripe images for robotic multi-layer multi-pass welding in this paper. Firstly, a conditional generator assisted by a weld pass number is established to guide the generator in processing images by introducing a prior information. At the same time, the above generator is streamlined through a lightweight convolution module to reduce the model scale. Then, a double-ended discriminator is constructed to assess the authenticity and class of the processed images, enhancing the processing effect through additional class loss. Finally, experiments show that the RMSE, PSNR, and SSIM of DARM are 1.384, 25.457, and 0.977, respectively, and the average processing time is 9.0 ms, which can achieve accurate real-time denoising and restoration of multi-layer multi-pass weld laser stripe images. The proposed method does not require image pairs for learning, the processing effect is better than that of existing methods, and the model scale is significantly reduced.