Automatic colorization using fully convolutional networks

Abstract

We propose an approach for automatically colorizing grayscale images using fully convolutional networks (FCNs). In contrast to traditional colorization methods, our approach operates only on grayscale images without any manual assistance. We first build an end-to-end deep learning network based on an FCN. Global, midlevel, and local features are extracted from the network and fused to construct each deconvolutional layer. To ensure color consistency, a low-frequency regularization term is presented to maintain the coherence between neighboring pixels. We then present an improved level-set method, which we apply to the output of the FCN to repair color bleeding caused by the rough segmentation performed by the FCN. To evaluate our approach, we compare the objective image quality resulting from our method with the results of other methods by assessing the peak signal-to-noise ratio, the mean squared error, the structural similarity index (SSIM), and the multiscale SSIM (MS-SSIM). In addition, we design a Turing test to evaluate the subjective image quality. The results show that our colorized images more closely resemble the ground-truth images and are more robust than those produced via other methods.