Natural Image Matting Based On Deep Learning Method

Digital image matting is one of the current hot research issues in computer vision,and it is widely used in computer special effects production,film and television works and other fields.The essence of image matting is the soft segmentation of images,aiming to extract foreground objects of interest to creators in pictures and video streams,and merge them with background stripping and new backgrounds to obtain new visually impactful pictures or video stream.Green screens are often used in the film industry to assist with matting,but in natural images,how to accurately extract foreground objects has become a research difficulty for researchers.The detailed information of the edge of the target object,including animal hair,semi-transparent objects,objects with similar colors,and blurred outlines will affect the accuracy of image matting to varying degrees.In order to solve the current problem of natural image matting accuracy,which is difficult to improve,based on the study of traditional matting methods based on propagation and color sampling,this paper proposes a natural image matting algorithm based on deep neural networks.The main contents are as follows:The various components of the classic convolutional neural network are studied,and the Res Net network is used as the backbone to improve.Different from the current single-shot encoding and decoding network commonly used in the field of image segmentation,it uses the U-Net structure.On this basis,the network in this paper uses multi-scale image features obtained from the original image,combined with the attention module of the channel domain,and learns low-level Image space and appearance information.In this way,the high-level opacity feature information is delivered to the feature multiplexing network that uses multiple encoding and decoding structures to increase the depth of the network and expand the receptive field on the convolution kernel.This method can improve spatial accuracy and strengthen the ability to acquire semantic information on the web.In addition,the feature dense connection module is added in the sampling process to efficiently transmit and reuse the same-level feature information encoded and decoded,and strengthen the network’s ability to extract non-neighbor features.The network of this article has been trained and tested on public data sets,and verified the matting effect and effectiveness of the different modules proposed in the article.Compared with the benchmark network,the experimental results show that the matting algorithm proposed in this paper can achieve higher accuracy.The test result based on the mean square error MSE on the Adobe data set is 25.7% higher than that of Deep Image Matting,which is an improvement in matting.The expected effect of graph accuracy.