The paper presents a fully automatic system for image colorization, leveraging deep neural networks to predict per-pixel color histograms. The approach integrates low-level and semantic representations, exploiting the fact that many scene elements naturally appear according to multimodal color distributions. The model is trained to predict hue and chroma distributions for each pixel, which can then be used to generate a color image or further manipulated. The system outperforms existing methods on both fully and partially automatic colorization tasks. The authors also explore colorization as a means of self-supervised visual representation learning, demonstrating that a colorization network trained from scratch on unlabeled images can achieve competitive performance on the Pascal VOC 2012 semantic segmentation task. The paper includes detailed experimental results, comparisons with prior work, and a new benchmark for automatic image colorization.The paper presents a fully automatic system for image colorization, leveraging deep neural networks to predict per-pixel color histograms. The approach integrates low-level and semantic representations, exploiting the fact that many scene elements naturally appear according to multimodal color distributions. The model is trained to predict hue and chroma distributions for each pixel, which can then be used to generate a color image or further manipulated. The system outperforms existing methods on both fully and partially automatic colorization tasks. The authors also explore colorization as a means of self-supervised visual representation learning, demonstrating that a colorization network trained from scratch on unlabeled images can achieve competitive performance on the Pascal VOC 2012 semantic segmentation task. The paper includes detailed experimental results, comparisons with prior work, and a new benchmark for automatic image colorization.