Jewelry Image-to-Image Translation with Consistency Regularization and Data Augmentations

Abstract

Image enhancement of jewelry is a difficult task because of the shape of the jewelry, its color, background elements such as shadows and glass stands, as well as the blurring of the boundary between the jewelry and the background and unique light reflections. Our preliminary results indicate that CycleGAN is effective in correcting jewelry images and that background elements in jewelry images adversely affect jewelry image correction. In this study, we propose a method to correct jewelry images with strong background elements. The results show that the target consistency of TC-ShadowGAN is effective not only in removing the background but also correcting the jewelry area in the image. In addition, data augmentation with Balanced Consistency Regularization (BCR) and Dense Consistency Regularization (DCR) are applied to increase the accuracy of the correction of the jewelry area.