| Recent advances in information computing and communication technologies have driven the development of new media technologies.As an important tool in the media production process,digital image processing technology has also gained a new development opportunity,especially the intelligent enabling technology for edge devices has further improved the accuracy of image processing tasks such as target detection,segmentation and reconstruction.However,as the technology continues to advance,the traditional hard segmentation scheme based on pixel classification model can no longer meet the demand for fine editing of short images/videos in the media era.As an alternative to soft segmentation based on regression model,Digital Image Matting(DIM)has become the hottest research topic in the field of image editing nowadays.The goal of digital matting is to capture the intended foreground targets from photos,audio and video streams,and blend them with new background materials to obtain new visually appealing images or video streams,while also serving to conserve captured materials.However,it is difficult to obtain better matting results with ideal computational overhead in existing digital matting.In order to balance the matting accuracy and computational power requirements as well as to correct the problem that natural images cannot accurately locate foreground targets,this thesis proposes a lightweight matting framework based on appearance cue guidance,and then fully automatic portrait matting is achieved by introducing contrast learning.The main contents of this thesis can be summarized as follows.1.This thesis designs a lightweight matting framework based on a multi-task structure,which divides the overall task into two types of subtasks.The first type of task is used to classify high-level features at the semantic level and to quickly divide foreground,background and unknown regions using high-level features;the other type of task is used to calculate the linear combination weights of unknown region features with respect to foreground and background layers.Accurate foreground features are obtained by sharing the weights of the high-level feature network with the feature classification task,and then fused with the low-level convolutional features.The proposed method is experimentally validated on the Adobe Composition-1K dataset and its synthetic dataset,and the results are compared with quantitative results to demonstrate the extremely fine matting masks obtained with low computational overhead.2.To address the problem of confusing the boundary semantics caused by unsupervised feature clustering of fully automated portrait matting algorithms,this thesis proposes a matting method based on twin network construction.To further increase the semantic gap between foreground and background features in the initial clustering,a contrast learning method is introduced in this thesis.Two isomorphic(twin)encoders are set to compute the correlation between detail features and semantic features in adjacent regions to explore the foreground and background target detail correlation information in the feature space from different image levels and assign the features to either the foreground network or the background network for iterative comparison training.The proposed method is experimentally validated on Alibaba PMM-100 and Adobe Composition-1K datasets,and the comparison of the results with quantitative results demonstrates that the proposed method is able to escape from the dependence of Trimaps on a large number of current matting algorithms.At the same time,the framework inherits the advantages of the previously studied lightweight framework,which can give finer portrait matting results with lower computational and data resource consumption in a fully automatic manner. |
