A Human Attribute Recognition Method Based on Attention Mechanism and Multi-task Learning

Abstract

The invention discloses a human body attribute recognition method based on attention mechanism and multi-task learning, which includes obtaining pedestrian images and processing them to obtain human body frames; constructing a shared convolution network to extract shared features for human body frames; The branch convolutional network of the branch convolution network uses the shared feature as the input of each branch convolutional network, and the output of each branch convolutional network is obtained as the personality feature of the corresponding attribute; the obtained personality characteristics of each attribute are respectively input into the corresponding attention of each attribute branch. The force mechanism network generates the attention map of each attribute, superimposes the attention map on the corresponding personality feature, and obtains the feature map of the region with the corresponding attribute; input the feature map to the fully connected layer corresponding to each attribute branch, and output Prediction and recognition results of various attributes of the human body. The present invention can learn the internal connection between each attribute, and obtain the key information area of ​​each attribute, so as to provide the accuracy rate of attribute recognition.

Description

technical field [0001] The present application belongs to the field of computer vision, and specifically relates to a human body attribute recognition method based on attention mechanism and multi-task learning. Background technique [0002] With the development of artificial intelligence and the large-scale deployment of high-definition video surveillance equipment, pedestrian attribute recognition has a good application prospect in video surveillance, smart retail, pedestrian re-identification and other fields, and has attracted more and more researchers' attention. And it has become a new research topic in the field of video surveillance systems. Video surveillance is distributed in every corner of the city. If effective information is extracted from a large amount of surveillance video information, it will inevitably consume a lot of manpower and material resources, and the efficiency will be low. Pedestrian attribute recognition is to extract the structural attributes ...

AI Technical Summary

Problems solved by technology

[0005] 1) For fine-grained attributes, such as glasses and jewelry, after multi-layer convolutional layer and pooling layer processing, the feature is weakened or disappears, and the traditional feature that directly extracts the entire image cannot identify these attributes well; in addition, Pedestrian attributes are different. Some attributes require shallow features, while some attributes require high-level features, some attributes require local features, and some attributes require global features to be recognized. How to extract a feature that can contain all the above for different attributes becomes a crucial issue

[0006] 2) The convergence speed of each attribute is different, which will cause different attributes to affect the recognition effect of other attributes during the training process

[0008] 4) The model training samples are all manually labeled, and pedestrians are all in the center of the frame. In practical applications, the input of attribute recognition is the detection output. Pedestrians may not be in the center of the pedestrian frame or the human body frame is incomplete, which affects the effect of attribute recognition.

Images