Lightweight image matting algorithm based on deep learning.

To solve the problem that the deep learning-based image matting algorithm cannot balance accuracy and model size, a lightweight image matting algorithm based on deep learning is proposed. Considering the limitation of memory and computing resources, and aiming at lightweight. We construct a network and gradually improved it. Firstly, apply deep detachable convolution to deep image matting networks to form faster and stronger encoder and decoder networks. The simultaneous use of depth-separable convolution can also reduce the number of corresponding model parameters and computation. And then attention mechanism is integrated into the model and SE Block was used to assign different weights to feature channels to improve the accuracy of the model. Finally, knowledge distillation scheme is designed in part of the encoder-decoder structure, the corresponding loss function is proposed, and the method of knowledge distillation is used to improve the feature learning ability of the lightweight neural network. Compared with the original deep image matting model, the number of parameters in the new model is greatly reduced without too much loss of accuracy.