An Interactive Method For Digital Image Matting Based On Recurrent Neural Network

Digital image matting is aimed to precisely pull out the foreground from the background in an image.It has been a hot research in the computer vision area for past decades.And the image matting technology also plays a key and fundamental role in image processing.In this digital media era,the booming of film,animation and television industry,together with the birth of some new applications such as webcast,make it more challenging for digital image matting.Meanwhile,deep learning technology,especially artificial neural networks can also provide new tools and aspects to solve image matting problem.In this work,the research therefore focuses on digital image matting problem,exploiting deep learning technology and artificial neural networks.Firstly,an interactive image matting framework based on recurrent neural network is presented.Taking advantages of the accurately-encoded image,the framework is able to iteratively suggest informative regions for matting in an active manner.Given these suggested regions,users only need to give a foreground or background label.These labels can automatically propagated to the whole image via the proposed framework,forming the final alpha mattes.Results demonstrate that the proposed framework can reduce users' efforts significantly and achieve comparable performance to the existing state-of-the-art image matting methods.Secondly,a densely-connected convolutional neural network is proposed to predict a trimap with rich semantics for an input image.The predicted trimap containing fore ground,background and unknown region has richer information than the saliency map generated by the salient object detection methods.The smaller unknown region area provided by trimap can lead a more accurate alpha matte estimation result for those trimap-based matting methods.Lastly,the proposed matting framework are concluded,their limitations are analyzed,and future work is put forward.