Decoloration of Multi-Color Fabric Images for Fabric Appearance Smoothness Evaluation by Supervised Image-to-Image Translation

For the poor adaptability of the two-dimensional image based fabric smoothness assessment methods for multi-color fabrics, three-dimensional imaging technologies were widely concerned in the area. In this paper, we suggest that the multi-color fabric smoothness assessment can be solved by the two-dimensional image based methods with the help of the proposed multi-color fabric decoloration method. The decoloration problem was solved by a paired image-to-image translation model built by conditional generative adversarial networks that was widely discussed in recent years. To train such model, we proposed a multi-color and white fabric image pairs generation method with a physically based fabric shading model. The experiment results show good performance of the proposed method on both generated image pairs and real multi-color fabric images. To quantitatively evaluate the performance of the decolored fabric images, a set of metrics was established based on the pixel wise image difference, fabric smoothness classification consistency, and manually fabric smoothness evaluation difference. The quantitative evaluation results demonstrated that the proposed method can achieve reasonable results for the multi-color fabric image decoloration.