Periodic Iterative Segmentation-Colorization Training: Line Drawing Colorization Using Text Tag with CBAMCat

Rui Luo, Yang Peng, Nuo Zhou, Jun Liang

Abstract: User-prompted line drawing colorization technology is gaining prominence for its potential to economize animation production. Nevertheless, traditional text-based coloring methods often grapple with issues such as color bleeding and errors due to the the absence of spatial and positional information in text tags. In response, we enhance training efficiency and effectiveness by periodically transitioning the network’s focus between segmentation and colorization stages. This innovation encourages the model to attend to crucial edge and shape features during colorization training, mitigating segmentation feature forgetting caused by prolonged training periods. Consequently, it effectively enhances coloring quality. In addition, we propose a novel network structure CBAMCat, which injects text tags information mapped to low-dimensional space into squeezed intermediate features, adaptively adjusts important features in the decoding block. This approach has proven effective in addressing localized coloring errors in confined areas. Qualitative and quantitative experiments substantiate the effectiveness of our proposed method.