Research On Portrait Style Transfer Algorithm Based On Convolutional Neural Network

As one of the important branches of image fusion,character style transfer aims to make one image have the information of multiple images at the same time,i.e.,it contains both the semantic information of content image and the style features of style image.However,the traditional person style transfer algorithm mainly has the following problems: when the background of the input person image is too complex and the proportion of faces is too low,it will lead to poor semantic segmentation,which results in the defects such as semantic style mismatch of the target image,distorted edge structure of the target image,distorted face structure,confused semantic content and poor image realism.Therefore,this paper conducts research on the style transfer method of portrait painting based on convolutional neural network,and achieves good artistic style transfer of portrait painting by image pre-processing of the input image and Deep Labv3+ model improvement.The main research is as follows:1.Content-aware clipping preprocessing research.For the problem that the face occupies a low proportion in the input image,which leads to the semantic segmentation cannot recognize the face details and the distribution of the five senses position better,the preprocessing is carried out before the semantic segmentation step,and the Las(Laplacian)operator edge detection is introduced to calculate the edges of the person image to obtain the edge information in the person portrait painting,and the color in the person image is used to find the region with higher saturation.Further filtering out the important regions in the person image.The face recognition procedure is also added to detect the proportion of the input portrait in the image and generate a target image with the face in the portrait painting as the main subject.Finally,the experiments show that the added content-aware cropping preprocessing method plays a good role in the semantic segmentation of the subsequent person images.2.Research on Deep Labv3+ network structure algorithm.To address the problems of semantic segmentation failure and segmentation discontinuity caused by distorted and deformed edge structure of face semantic segmentation and unclear recognition of five senses,the Deep Labv3+ model is improved,firstly,the backbone network is a lightweight convolutional network Efficient Netv2 network,and the utilization of model parameters is effectively optimized by training perceptual neural architecture search and scaling techniques to improve training efficiency.At the same time,the global average pooling operation is improved by replacing it with a mixed strip pooling operation,which helps to capture the details of people such as hair,facial features and clothes while taking into account the global and local information of the image,and improves the complexity of the model in processing people images.The experiments show that the improved Deep Labv3+ network structure algorithm effectively optimizes the accuracy and generalization performance of the model semantic segmentation,and achieves better experimental results for semantic segmentation of human images.