MccSTN: Attention and Multiscale Contrast Based Style Transfer of Artistic Images

Abstract In the realm of image processing today, creative image alteration plays a significant role. Recent work on the subject of style transfer has employed attention processes to access the aesthetic-artistic expression of style images. This method converts style images into tokens by first determining the level of attention and then using a decoder to transfer the image’s artistic style. Nevertheless, this leads to the discarding of many fine-grained style elements because of the extremely low semantic similarity between the original image and the style image. Discordant or blatant artifacts may result from this. We suggest MccSTN, a novel framework for style representation and transfer that may be tailored to work with current arbitrary picture style transfers, as a solution to this issue. To be more precise, we first present the Mccformer feature fusion module, which fuses fine-grained features from content photos with aesthetic characteristics from style images. Mccformer is used to obtain feature maps. The desired image is subsequently obtained by feeding the feature map into the decoder. We take into account the relationship between individual styles and the general style distribution in order to lighten the model and train it quickly. We describe a multi-scale augmented contrast module that uses a vast number of image pairs to learn style representations. Code will be posted on https://github.com/haizhu12/MccSTN