Research on the expansion method of metal wear sample library based on improved CycleGAN network

Metal wear detection is an essential aspect of metal materials research. However, obtaining wear data for certain metals can be challenging, impacting detection accuracy. This paper proposes expanding the metal wear sample library using an improved CycleGAN. This method optimizes the composite model by replacing the L2 paradigm of the cycle loss function with the LPIPS function. This enables the generated images to have the fundamental outlines of metal wear samples and additional detail variability while enhancing the training network's resolution and stability. Experimental results demonstrate that the CycleGAN network with the improved loss function significantly improves the quality of the generated images from the metal wear dataset compared to the primary CycleGAN network. Specifically, the SSIM and PSNR values enhanced by 6.38% and 9.34%, respectively.