Graph convolutional network-based image matting algorithm for computer vision applications

Image matting plays a vital role in a variety of computer vision tasks including video editing and image fusion. Previously presented image matting algorithms might fail in producing favorable results since most of them concentrate on the similarity between the neighboring pixels while neglecting the corresponding spatial relationship. To address this issue, an end-to-end image matting framework through leveraging deep learning mechanism and graph theory is proposed. The proposed pipeline is a concatenation of one deep feature extraction component and a Graph Convolutional Network (GCN). The former part takes an image and its corresponding trimap as inputs and can generate the pixel-wise features, which are then exploited as the input of the GCN locating at the latter part of the proposed framework. The GCN would refine the features for every pixel and predict the alpha matte outcome of the image. The approach outperforms a group of state-of-the-art matting techniques as shown by the theoretical analysis and experimental results in terms of both accuracy and visual effects.