Abstract:

Video colourisation methods that transfer colour information in keyframes based on traditional optical flow are time-consuming, while those relying on global colour transfer are prone to desaturation. This paper proposes a new video colourisation method based on voxel flow. In the proposed method, the reference and target images are both converted to the lab colour space, before a double-channel voxel flow is obtained by feeding the luminance channels of the images into a neural network. The voxel flow values indicate the positional colour correspondence between the target frame and the reference frame. Then, the colour of the target frame is obtained by bilinear interpolation of the reference frame utilising the voxel flow. Finally, the colour and luminance channels are combined to synthesise the final colourised image. Experimental results show that the proposed video colourisation method maintains the saturation of the reference image, while also maintaining edge sharpness. Compared with rival video colourisation methods based on traditional optical flow, the proposed method yields a higher peak signal-to-noise ratio (PSNR) and offers a shorter runtime.

Key words: colourisation, voxel flow, deep learning, neural network, optical flow