No abstract

No abstract

Recent progress on image colorization is substantial and benefiting mostly from the great development of the deep convolutional neural networks. However, one type of object can be colored by different kinds of colors. Due to the uncertain relationship between the object and color, the deep neural network is unstable and difficult to converge during the training process. In order to solve this problem, this paper proposes an instance-aware automatic image colorization algorithm, which uses the semantic features of the object instance as prior knowledge to guide the deep neural network to do the colorization task. Meanwhile, we design a discrete loss function to train the deep network and this network can be trained from end to end. Experiments show that this algorithm can obtain satisfactory colorful results on the images containing object instance and achieves state-of-the-art results.

No abstract

No abstract