Design And Implementation Of Image Style Transfer Method Based On Deep Learning

With the continuous progress of in-depth learning,it is found that in-depth learning can be used to train models for various scenario applications,such as biometrics.The extraction of useful features is an important step in object recognition.Traditional non-in-depth learning methods do not have universality for image texture processing,which usually uses mathematical methods to build specific models.Neural networks in deep learning can automatically extract the most useful features after training.Instead of manual segmentation using traditional modeling methods,they can automatically select the optimal values through neural networks.Image style transfer is about the extraction of a content map and a style feature and texture synthesis.The method of texture synthesis and style transfer using CNN can produce good experimental results.This paper will point out some limitations of these methods in texture quality,stability and parameter adjustment,and propose a multi-control texture synthesis method based on cosine similarity constraints of image features to improve the shortcomings of traditional methods.Firstly,this paper gives an explanation of the mathematical principle for the instability of the original problem,and then uses the cosine similarity of feature blocks to match the feature blocks of th e style image and the content image to improve the instability.This paper also shows how to adjust the local style loss in the multi-scale framework of this paper.The method proposed in this paper can improve the quality in very few iterations and is mor e stable in optimization.This method can improve quality,improve the distinction between content and style,and provide artistic control.Based on TensorFlow platform,this paper presents a local matching method based on cosine similarity constraints for image style transfer.Compared with previous methods,the whole result not only achieves excellent effect on image style transfer,but also supports arbitrary input of content and style,which makes the whole result more flexible and humane.