Generation Model of Gender-forged Face Image Based on Improved CycleGAN

Deep forgery can stitch people's voices, faces, and body movements to synthesize false content for gender conversion, age change, etc. Face gender forgery images based on generative adversarial image translation networks are prone to change irrelevant image domains, face Insufficient details and other problems. In response to these problems, this paper proposes a face gender fake image generation model based on improved Cy-cleGAN. First, the generator structure is optimized, and the attention mechanism and adaptive residual block are used to extract richer facial features. ; Then, the loss function is improved by referring to the idea of ​​relative loss to improve the discriminant ability. Finally, a model training strategy based on age constraints is proposed to reduce the influence of age changes on the generated images. In CelebA and IMDB-WIKI data The experimental results show that, compared with the original CycleGAN method and the UGATIT method, the proposed method can generate more realistic face gende