Automatic ID Photos Matting Based on Improved CNN
Bin Qian, Xiguang Gu · 2019
Image matting that aims to estimate the foreground from the mixed pixels is key to ID photos processing. Recently, deep convolutional neural network (CNN) has shown competitiveness for image segmentation, but not achieved good accuracy in practical matting. In this paper, we propose an improved light network for ID photos matting. The model adopts the MobileNetV2 as backbone and leverages the densely connected blocks to predict a high-accuracy binary mask for K-nearest neighbor (KNN) matting. In addition, an automatic portrait matting system based on L-UNet is built, which is very fast without any interaction. Experimental results on a large scale real-world dataset verify the effectiveness of the proposed method on ID photos matting task.