Inverible Gray Image with Predictive Coding

The invertible grayscale technique possesses excellent color restorability by implicitly encoding color information. However, the color-encoded texture patterns are susceptible to external image manipulations, such as JPEG compression, which limits their practical applications. One natural approach is to introduce disturbances during training to learn a robust encoding scheme. Nevertheless, it is challenging to maintain both grayscale visual quality and the robustness of color restorability in the encoding scheme. This means that the color-encoded texture patterns may become amplified. To address this, we propose a novel approach inspired by predictive coding, where we utilize a pre-trained colorization model as a decoder to expand the encoding space. This way, only distinctive color information needs to be encoded. Our experimental results demonstrate that our method can generate noise-tolerant invertible grays cales without compromising visual quality. Furthermore, a comparison with feasible baselines validates the superiority of our proposed designs.