Application of Convolution Neural Network in Landscape Painting Style Transfer and Feature Extraction Technique

Abstract

The study of style transfer and feature extraction in computer-based painting is a difficult topic. The majority of research employs manual segmentation to select local regions, leading to ineffective final extracted features and an inadequate capacity to discern the aesthetic approach of the artist. This research proposes a unique style loss and content loss-guided two-channel VGG network for feature extraction and landscape painting style transfer, aiming to address this issue. After convolving the input image, we employ the feature maps across the two-channel network's third through fifth layers. To achieve feature extraction, feature constraint addition, and parameter control for traditional Chinese paintings with local style transfer, we construct a loss function for content and style from the higher layers to the lower layers, which a decoder then decodes to the next layer after each layer matches. We continue this process until we obtain the final synthetic map. We observe significant differences in feature information, such as lines, textures, and frequencies, between landscape paintings and other painting styles. As a result, it is possible to extract, constrain, and learn this feature information while losing style and content.