Abstract:

Human parsing, as a complex and delicate computer vision task, has a wide application prospect. In order to obtain accurate human parsing results, it is necessary to extract rich human semantic features. In this paper, a two-way human parsing network（MTCnet） is proposed. MTCnet combines the encoding and decoding network with atrous convolution, it has two feature extraction subnets that can integrate the learning of multi-scale feature information. Compared with the single network, it can learn richer semantic feature information of human. Different from the previous one-step approach, the method proposed in this paper needs to carry out multi-stage learning. Each stage improves the human parsing results of the previous stage, and finally achieves the optimal result. The experimental results show that the proposed method has stronger feature extraction ability and more accurate human parsing results than some advanced methods.

Key words: human parsing, atrous convolution, encoder-decoder network, deep learning

摘要：

人体解析作为一种复杂而精细的计算机视觉任务，应用前景十分广泛，为了得到精确的人体解析结果需要提取丰富人体语义特征，对此提出一种双路人体解析网络（MTCnet）。MTCnet将编码解码网络与空洞卷积相结合，拥有两个特征提取子网能够融合学习多尺度特征信息，相比单一网络，能够学习更丰富人体语义特征信息。与以往单一阶段处理方式不同，提出的方法需要进行多阶段学习，每个阶段对前一阶段的人体解析结果进行改进，达到最优的解析结果。实验结果表明，提出的方法与目前一些先进的方法相比特征提取能力更强，解析结果更加精准。

关键词: 人体解析, 空洞卷积, 编码解码网络, 深度学习