Improving Reference-Based Image Colorization For Line Arts Via Feature Aggregation And Contrastive Learning.

The tremendous semantic discrepancy between the line art drawings without texture and the reference pictures containing rich color challenges current image-to-image translation models. Previous works attempt to establish cross-domain correspondence. However, they fail to capture more detailed features. A Reference-based Line art Translation Network (RLTN) is introduced with a Multi-level Feature Aggregation Module (MFAM) to improve the performance. The MFAM concentrates on more meaningful information for feature matching by utilizing the Multi-stream High Frequency Block (MHFB) and the Pixel-wise Correlation Block (PCB). We also employ the Channel-level Attention Block (CAB) and the Spatial-level Attention Block (SAB) for a better fusion of features. Moreover, a Style-based Contrastive Loss (SCL) is proposed to maintain the style similarity between the synthesized images and the reference examples. Experiments conducted on three datasets demonstrate the effectiveness of our model in producing more pleasing visual effects compared with state-of-the-art approaches.