Multi-Feature Fusion Network for Image Artistic Style Transfer

Image artistic style transfer technology has become a crucial tool in fields such as scene rendering and game design. However, existing methods often struggle to effectively capture local texture details during the feature extraction phase. This leads to inconsistent styles in the generated images and makes it difficult to balance the integrity of the content structure in complex scenes. To address these issues, this paper proposes a multi-feature fusion network. First, this network achieves the extraction from rough artistic features to refined representations through the collaborative work of a deep convolutional neural network and an image encoder. Then, it uses the feature map structure constructed by the 1-nearest neighbor algorithm to enhance the associations between features. Moreover, a dual-branch up-sampling and down-sampling structure guided by cross-entropy loss and mean-squared error loss, respectively, is employed to achieve the adaptive fusion of the content structure and the artistic style. During this process, the spectral filter, as the core module, further optimizes the quality of the fused feature map. Finally, the fused high-dimensional features are end-to-end mapped to high-quality stylized images via a generative adversarial network. Experimental results show that the proposed method achieves average values of 0.72 in structural similarity and 23.8 in peak signal-to-noise ratio, respectively. The performance indicated by these values substantially outperforms that of existing mainstream methods, demonstrating its good performance in generating visual quality and style consistency, as well as its potential for practical applications.