IMAGE STYLE TRANSFER MODEL WITH CONTENT PRESERVATION
Ruixue Liu, Shengbei Wang, Weitao Yuan · International Educational Applied Scientific Research Journal · 2019
The image rendered in another image style, called style transfer, is one of the most interesting applications of deep learning. Style transfer plays an increasingly important role in basic computer vision research and industrial applications. The method of iterative optimization [1] and training feedforward convolutional neural network [2-4] are used to continuously achieve fast and good effects on style transfer. However, most of these methods ignore an essential problem that the content and contour information of the original images could be lost after the transfer process. Throughout the history of style transfer, although the performance is constantly improving, pursuing higher flexibility and speed, an essential problem is ignored, that is, the main content and contour are unavoidably blurred after style transfer. It is found that in many application fields, the content/contour of the original image has great importance and thus should be kept as clear as possible after style transfer. In this paper, a sub-structure named Important Content Contour Extraction (ICCE) is proposed to generate masks and therefore preserve the clear contour and content after style transfer. The Feature Extraction Module (FEM) for extracting advanced features is added in image style transfer network which is trained by the perceptual loss function. We conducted some experiments to verify the effectiveness of the proposed method. Experiments demonstrated that the proposed framework achieves a good compromise in speed, flexibility and quality. The experiment results showed that the proposed model had good ability for style transfer.Meanwhile, it could preserve clear content and contour of the original images, which was better than [1-4].