Image Realism Enhancement Method Based on Improved Neural Neighbor Style Transfer

Object detection technology is a very important technique that relies on a large amount of data for training. However, obtaining sufficient data for specific detection targets can be challenging. In cases where data are limited, it is necessary to increase the number of datasets using various methods. We propose an improved algorithm based on the Neural Neighbor Style Transfer (NNST) algorithm to better adapt to the task of realistic style transfer. Through comparative experiments, we find that image realism could be enhanced when the alpha value is 0.75 and content loss, color correction, and high pixel output are used. Additionally, we design a more effective feature extraction network called SE-VGG19. Compared to the original VGG16, SE-VGG19 can improve the network's ability to perceive style and extract features, making the generated images match the target style better while preserving the original content features. Furthermore, we suggest using the center cosine distance instead of the original Euclidean distance for loss measurement. After comparison and verification, our method has been proven to improve image realism compared to the original algorithm greatly.