A Lightweight Generative Adversarial Network for Artistic Style Transfer
Pingan Qiao, Tongtong He · 2024
Currently, image style transfer is one of the hotspots in image processing. Generative adversarial networks have been successfully applied in many fields by virtue of their advantages, and the technology of image style transfer has also seen improvements. However, there are still some problems with the existing artistic style transfer models. For example, the generated images will contain noisy textures, the style and color effects are not ideal, and the model has many parameters, which is easy to cause model breakdown and unstable training. To overcome these problems, an image transfer method that improves the cycleconsistent generative adversarial network (CycleGAN) is presented in this paper. According to its structural characteristics, distributed shift convolution (DSConv) is introduced through the quantized neural network, and the reverse residual block is used to optimize the generator network structure, which can greatly reduce the calculation amount of the model while also improving the model's perception ability and accuracy and the quality of the generated images. Experiments show that the proposed approach can improve the content details of the generated image andthe color effect of the style texture, significantly reduces the number of model parameters, reduces the computational cost, improves the transfer rate, upgrading the model performance, and has strong style transfer capabilities.