Image Colorization with Dense Feature Fusion

We propose a new model for colorizing grayscale images with a U-Net-like network structure that focus on the connection between global and local features. A novel skip connection method is adopted to change the way information flows, which incorporating multi-scale feature information. This enables us to obtain more common features of encoding and decoding layers. Low-level detail features and high-level location features are exactly the semantic information we need. We argue that these semantic information plays an important role in the model's learning of colorization tasks. When there is as much similar semantic information as possible from the decoder and encoder networks, the network will handle easier learning tasks. The proposed model architecture is evaluated on a large dataset for gray image colorization. Experimental results show that our model improve the coloring performance.