Research on Image Style Transfer Technique Based on Normalization
Hanmin Ye, Zhibo Li, Wenjie Liu · 2021 IEEE 5th Information Technology,Networking,Electronic and Automation Control Conference (ITNEC) · 2021
In this paper, the existing Gram algorithm based on CNN network is improved to calculate the image features of mean standard deviation. In this method, the feature space is constructed to store the feature information of different filters, and different normalized statistics are realized in different network layers, so that multi-scale and stable features can be obtained better. It does not need to train the real data, and style transfer can be carried out flexibly. According to the theoretical analysis of CNN, the feature information of the high-level network reconstruction is constructed and extracted, and the FC layer and soft-max layer are removed to improve the operating efficiency. Experimental results show that the performance of the proposed mean-standard deviation algorithm is better than that of the Gram algorithm in the process of style transfer. Meanwhile, the distortion effect of style transfer is small, and the running efficiency is significantly improved.