Abstract: Aiming at the problems of low completeness of character matting, insufficiently refined edges, and
cumbersome matting in matting tasks, an automatic matting algorithm for portraits based on deep learning was
proposed. The algorithm employed a three-branch network for learning: the semantic information of the
semantic segmentation branch (SSB) learning  graph, and the detailed information of the detail branch (DB)
learning  graph. The combination branch (COM) summarized the learning results of the two branches. First, the
algorithm’s coding network utilized a lightweight convolutional neural network MobileNetV2, aiming to
accelerate the feature extraction process of the algorithm. Second, an attention mechanism was added to the SSB
branch to weight the importance of image feature channels, the atrous spatial pyramid pooling module was added
to the DB branch, and multi-scale fusion was achieved for the features extracted from the different receptive
fields of the image. Then, the two branches of the decoding network merged the features extracted by the
encoding network at different stages through the jump connection, thus conducting the decoding. Finally, the
features learned by the two branches were fused together to obtain the image  graph. The experimental results
show that on the public data set, this algorithm can outperform the semi-automatic and fully automatic matting algorithms based on deep learning, and that the effect of real-time streaming video matting is superior to that of
Modnet.

Key words: fully automatic matting, lightweight convolutional neural network, attention mechanism, atrous spatial
pyramid pooling, feature fusion