Multi-semantic preserving neural style transfer based on Y channel information of image

Neural style transfer, as a new auxiliary means for digital art design, can reduce the threshold of technical design and improve the efficiency of creation. The existing methods have achieved good results in terms of speed and style quantity, but most of them change or erase the semantic information of the original content image to varying degrees during the process of stylization, resulting in the loss of most of the original content features and emotion; although some methods can maintain specific original semantic mentioned above, they need to introduce a corresponding semantic description network, leading to a relatively complex stylization framework. In this paper, we propose a multi-semantic preserving fast style transfer approach based on Y channel information. By constructing a multi-semantic loss consisting of a feature loss and a structure loss derived from a pre-trained VGG network with the input of Y channel image and content image, the training of stylization model is constrained to realize the multi-semantic preservation. The experiments indicate that our stylization model is relatively light and simple, and the generated artworks can effectively maintain the original multi-semantic information including salience, depth and edge semantics, emphasize the original content features and emotional expression and show better visual effects.