ECM: arbitrary style transfer via Enhanced-Channel Module

Arbitrary style transfer is to fuse the style of one image with the content of another image. With the development of artificial intelligence, many style transfer methods have emerged, which focus on stylizing content features in different aspects, such as loss functions, attention mechanism, and functional instance normalization layer. However, there are still problems such as the lack of local detail information or the insufficient degree of style fusion in the generated images. To get more realistic images, we propose an Enhanced-Channel Module, which vectorizes content feature maps and style feature maps to generate content-aware channel weights. The channel weights are multiplied by the content feature maps as the map offsets to complete the style injection. Meanwhile, in order to utilize the multilayer feature maps, we train the network in a progressive rendering way using multilayer content features and deep style features. The comparison experiments with the state-of-the-art methods prove that the proposed method can generate high-quality stylization results in artistic style transfer and video style transfer tasks, and the ablation experiments prove the effectiveness of the different components.