Lightweight image matting algorithm based on deep learning

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