Human body analysis method based on graph representation and improved Transform

Abstract

The invention discloses a human body analysis method based on graph representation and an improved Transform. According to the method, high-dimensional feature representation is embedded into low-dimensional graph features, the improved Transform is used for reasoning calculation and capturing context feature relations, new graph features are generated and decoded again to form a fine analysis graph, and therefore the whole model is iteratively trained in an efficient mode to obtain a final analysis result. According to the method, reasoning calculation is performed more efficiently only according to priori knowledge of a human body hierarchical structure; reasoning is performed on the human body part features represented by the graph, so that more calculation cost can be saved in subsequent iterative reasoning; and the structure of Transform is improved, and the context information of the features of each part of the human body is extracted and integrated globally, so that the association degree of different parts of the human body is perceived comprehensively, and the precision of an analysis result is higher.

Description

technical field [0001] The invention relates to the technical field of computer vision and image processing, in particular to a human body analysis method based on graph representation and improved Transformer. Background technique [0002] Human body parsing is an important and challenging topic in computer vision. It understands various parts of the human body by intensively predicting each pixel, thereby dividing multiple semantics. Recent studies have shown that human body parsing is widely used in human body analysis tasks, such as human body image generation, virtual fitting, pose estimation, pedestrian re-identification, etc. [0003] Human body parsing is the semantic segmentation of the human body, which actually completes the pixel-level classification. Since each pixel corresponds to a different semantic label, these different categories of semantic labels have shared features and certain relevance. The current method is divided into three aspects: one is to use...

AI Technical Summary

Problems solved by technology

[0005] The disadvantage of this method is that using pose and edge features requires a lot of additional prior data for training

[0007] The disadvantage of this method is that the three kinds of feature maps are used as the input Transformer sequence, and the correlation degree is calculated pixel by pixel, which consumes a lot of computing resources.

[0009] The disadvantage of this method is: the graph convolutional neural network of this method needs to iterate continuously to make up for the shortcomings of insufficient global information correlation, because in the process of message passing mechanism, the shared information between different human body structure levels is difficult to calculate in parallel, Need to wait for the graph node information to be updated one by one

Images