TFMNet: Trimap-free real-time image matting algorithm based on deep learning

Abstract

The conventional image matting algorithms needed priori manual Trimap information to produce excellent matting results which made real time matting impossible. To tackle the problem, a Trimap-free image matting network, TFMNet, is proposed in this paper. The proposed network consists of four modules, ConvNeXt backbone module for image features extraction, Trimap prediction module for normalized Trimap generation, glance matting module for rough matting results prediction, and post-processing module for exact matting results production. To further optimize the training process of the proposed model, an improved Loss function based on frequency domain information is proposed. In experiment, Sets of Experiments designed by variable controlling approach prove that the proposed TFMNet do well in real time image matting. The TFMNet model achieves 8.99, 0.011, 12.31, 11.15 in the accuracy metrics of SAD, MSE, GRAD, CONN, respectively, costs 51ms for one image averagely which meet the real-time requirements, and model size is 671M. Besides, further experiments conducted by comparing with five state-of-the-art models based on three typical matting databases demonstrate the superiority of the proposed algorithm.

Keywords

• real-time image matting
• image semantic segmentation
• convolution neural network without pooling
• image processing