Invertible Grayscale via Dual Features Ensemble

Abstract

Grayscale image colorization is known as an ill-posed problem because of the imbalanced matching between intensity and color values. Even given prior hints about the original color image, existing colorization methods cannot recover the original color image from grayscale faithfully. In this paper, we propose to embed color information into an invertible grayscale, such that it can be easily recovered to the original color. However, a vanilla encoding-decoding network cannot produce rich representations of color information and thus the reconstruction quality is limited. Moreover, due to the neglect of the discrimination of color information, it cannot embed color information into visually inconspicuous patterns located in the grayscale. In this paper, we propose a novel color-encoding schema, dual features ensemble network (DFENet), for the effective embedding and faithfully reconstruction. In particular, we complement the residual representations with dense representations, to integrate the ability of local residual learning and local feature fusion. Furthermore, we propose an element-wise self-attention mechanism that highlights the discriminative features and suppresses the redundant ones generated from the dual path module. Extensive experiments demonstrate the proposed method outperforms state-of-the-art methods in terms of reconstruction quality as well as the similarity between the generated invertible grayscale and its groundtruth.