Semantically Aware Style-Controlled Animation Line Art Colorization Using Conditional GANs with GCN and Attention Mechanisms

Abstract

In traditional animation production, coloring line art is a labor-intensive and time-consuming process. In this study, image processing techniques based on generative adversarial networks (GANs) were investigated to develop an algorithm for the automatic coloring of animation line drawings. The objective was to improve production efficiency, reduce the workload of artists, and generate outputs that are natural in appearance, consistent in style, and closed along region boundaries. The proposed method was implemented within the conditional GAN framework. The generator adopted a U-Net architecture with skip connections, which allowed both fine-grained details and global structural features of the line art to be captured. A spectrally normalized discriminator was used to evaluate the realism of local image regions. To improve semantic accuracy and color coherence, an attention mechanism was incorporated, enabling the model to focus on key semantic areas and learn dependencies between color regions. End-to-end training was conducted using a large-scale paired dataset of line art and corresponding colored images. A multi-task learning strategy combining perceptual loss, L1 loss, and adversarial loss was employed for optimization. Latent space interpolation was further introduced to allow limited user adjustment of color styles. Experimental results indicated that the algorithm achieved a PSNR of 30.2 dB and an SSIM of 0.94, which represented improvements of 2.2 dB and 0.04 over the Structure Probe Adaptive Differential Evolution (SPADE) baseline, respectively. The FID and boundary overflow rate were reduced to 18.3 and 2.1%, also showing clear improvements over the baseline. With the inclusion of structural consistency loss, graph convolutional networks, and self-attention mechanisms, the method maintained accurate boundary preservation, achieved cross-region color consistency, and supported adaptive style rendering. In summary, the algorithm addressed key limitations of existing approaches, such as color overflow, semantic inconsistency, and rigid stylization. These findings demonstrate improvements in automatic coloring quality and suggest the potential of GAN-based techniques in animation production.