Research On Image Style Transformation Algorithm Based On Deep Learning

Image style transformation is a technology that can transfer the texture information of a particular artistic style image to another natural image,and ensure that the original natural image has the texture feature of the specific artistic style while maintaining the semantic content basically unchanged.Nowadays,deep learning with its excellent learning and image processing ability has become the most popular direction in the field of computer vision.Based on the deep learning method,this paper does some research on image style transformation algorithm.This paper first introduces the research value of current style transformation algorithms in computer vision applications,domestic and foreign research results and development trends.The limitations of traditional algorithms are explained,the key technologies of various current deep learning-based image style transformation algorithms are introduced,their respective advantages and disadvantages and their application scenarios are compared.Finally,a multi-scale feature fusion network designed in this paper is introduced.This network fuses multi-scale feature information to complete the style transformation task,and achieves a better style transformation effect in details.The multi-scale feature fusion style transformation algorithm designed in this paper is mainly for the inevitable loss of some local information in deep-level convolutions,so we consider weighting decoding of convolution results at different levels.The specific work is as follows:?1?The encoder network consisting of a convolutional neural network is used to obtain content feature maps and style feature maps of different scales.The encoder structure uses the pre-trained network VGG from conv1₁ to conv4₁.The features of the content and style feature map at conv2₁,conv3₁,and conv4₁ are activated for all outputs,preparing for multi-scale feature fusion.?2?Use the whitening coloring transformation algorithm to perform feature transformation on the acquired content feature maps and style feature maps at different scales,match the content feature map distribution with the style feature map,and get fusion feature maps of different scales.In order to prevent the influence of noise on the feature transformation in the content and style feature maps,singular value decomposition algorithm is used to remove the information with smaller feature values,in this way,the influence of the noise that may exist on the image feature transformation is avoided while not affecting the overall characteristics of the image.?3?The fusion feature maps of different scales are weighted from different positions into the decoder network designed by this paper.The decoder finally restores the multiscale fusion feature map back to the original image space to complete the style transformation.