Colorful Image Colorization

摘要: Given a grayscale photograph as input, this paper attacks the problem of hallucinating plausible color version photograph. This is clearly underconstrained, so previous approaches have either relied on significant user interaction or resulted in desaturated colorizations. We propose fully automatic approach that produces vibrant and realistic embrace underlying uncertainty by posing it classification task use class-rebalancing at training time to increase diversity colors result. The system implemented feed-forward pass CNN test trained over million images. evaluate our algorithm using “colorization Turing test,” asking human participants choose between generated ground truth image. Our method successfully fools humans 32 % trials, significantly higher than methods. Moreover, we show colorization can be powerful pretext for self-supervised feature learning, acting cross-channel encoder. results state-of-the-art performance several learning benchmarks.