A Novel Human Parsing Method Driven by Multi-Scale Feature Blend Network

In recent years, human parsing has been developed a lot for its valuable utilization. However, existing methods have not fully solved semantic errors and incomplete semantic predictions. In this regard, a Multi-Scale Feature Blend Network(MFBNet) is proposed to deal with these problems from the respective of fusing multi-scale features. Specifically, we creatively introduce the Context Embedding module which uses the feature pyramid as the main structure to blend multi-scale feature information. Besides, ResNet-101 is applied as the backbone network to train and optimize shared weights and map the generated feature maps to the Context Embedding module. Experimental results on several wide-used datasets show that the proposed method outperforms than the state-of-art methods in human parsing.