Generative adversarial network based on CNN classifier predicted scores for image style transfer

In recent years, image style transfer has emerged as an increasingly popular research theme in the field of computer vision. Image style transfer seeks to convert the style of an image or mix it with other styles to produce an image with stylistic features not found in the original image. In the past, image style conversion was mainly implemented using feature conversion or filters, processes which require a significant degree of manual design, thus limiting these approaches to performing at most a single style conversion. However, with the development of deep learning technologies, an increasing number of studies have begun to break through these research challenges, achieving significant progress. The study proposes generating image data attribute tags from the classification dataset, constructing a GAN architecture for multi-style conversion tasks, and proposes a Classifier Style Scores GAN (CSS-GAN) model. First, a CNN classifier is used to train on the classification dataset. Once its stability is verified, the classifier is used to make predictions and its output layer features are extracted as attribute labels. These labels are subjected to different types of pre-processing to assess the performance difference between smoothed labels and binary classification labels. Finally, the resulting labels are used to train a multi-style transfer GAN. Experiments are conducted using a facial attribute dataset to compare the labeling method with the proposed model architecture. The results indicate that using classifier-predicted features and applying feature smoothing as attribute labels for training the GAN can effectively enhance the quality and stability of images generated for the style transfer task. Additionally, this approach allows for better control over the degree of transformation and improves the overall performance of style transfer.