High-Precision Anime Conversion Model Based on Generative Adversarial Networks

Currently, the application of anime image conversion is becoming increasingly widespread. However, in the task of converting real images to anime images, traditional convolution operations can lead to information loss and blurring. Therefore, there are problems such as unstable network training, severe distortion of generated images, and blurring of generated images. This paper proposes an improved model GI_CartoonGAN(Group convolution channel shuffle and Inception dilated convolution Cartoon Generative Adversarial Network) used for anime image conversion. On the basis of the CartoonGAN model, this network model improves the representational capabilities of the generated network by introducing grouping convolution channel shuffle operations, enriches image features, improves the accuracy and expressiveness of feature extraction, improves image conversion accuracy, and introduces the concept structure and dilation convolution operation to expand the receptive field of the convolution kernel, effectively processing features at various scales of the image, Improve the ability of generator to model features and perceive different styles and details, thereby enhancing its ability to generate detailed features. The experimental results show that the FID index of the image generated by this model has an improvement of over 17% compared to other models, which can effectively improve the clarity and authenticity of the generated images.