An Improved Deep Convolutional Generative Adversarial Network for Anime-Style Character Image Painting

The application of Deep Convolutional Generative Adversarial Networks (DCGANs) has gained significant popularity in the domain of anime-style image painting. However, the model often suffers from the imbalanced training process and other certain limitations, such as the presence of blurred regions and feature misalignment. In our work, an enhanced DCGAN is proposed to tackle these challenges. First of all, we introduce a scoring mechanism in each convolutional layer of the discriminator to enhance its ability to capture multi-scale features for accurate rating. From another perspective, we integrate the self-attention mechanism and modify the sizes of convolutional kernels in the generator, resulting in greater capacity to generate details. To ensure training stability, we employ the WGAN-GP loss, which also improves the generation diversity. Furthermore, in order to provide a fair and thorough evaluation of the generated outputs, we employ a comprehensive human-machine scoring method to evaluate the outcomes. The experimental results demonstrate that our improved DCGAN exhibits stronger feature extraction capabilities and effectively enhances the quality of generated images.