Face attribute recognition method based on multi-task multi-label learning convolutional neural network

Abstract

The invention discloses a face attribute recognition method based on a multi-task multi-label learning convolutional neural network, and relates to a computer vision technology. Firstly, multi-task learning is adopted, and two tasks of face key point detection and face attribute recognition are learned at the same time; different learning difficulties and learning convergence rates of different attributes are considered, the attributes are divided into subjective attributes and objective attributes, and the convergence rate of the network is accelerated and the problem of sample imbalance is relieved by adopting a dynamic weight and an adaptive threshold strategy; and finally, according to the trained network model, face attribute recognition results of the subjective attribute and objective attribute sub-networks are taken as final face attribute recognition results. A dynamic weight scheme and self-adaptive threshold adjustment are used, so that the network convergence speed is increased, and meanwhile, the label imbalance problem can be relieved; three different sub-networks are trained by adopting a spatial pyramid pooling method, so that end-to-end training is achieved to perform multi-task multi-face attribute recognition. And the precision of face attribute recognition, especially subjective attributes with high difficulty, is improved.

Description

technical field [0001] The invention relates to computer vision technology, in particular to a face attribute recognition method based on multi-task and multi-label learning convolutional neural network. Background technique [0002] In the past few years, facial attribute recognition has attracted great attention in computer vision and pattern recognition, with major applications including image retrieval, face recognition, pedestrian re-identification, micro-expression recognition, image generation, and recommendation systems. The task of face attribute recognition is: given a face image, predict multiple face attributes; such as gender, attractiveness, and smile. Although the task of face attribute recognition is only an image-level classification task, there are still many challenges, mainly due to changes in facial expressions caused by changes in face angles and lighting. [0003] Recently, due to the outstanding performance of Convolutional Neural Networks (CNNs), me...

AI Technical Summary

Problems solved by technology

But these methods treat various attributes equally (using the same network architecture for all attributes), regardless of the different learning complexities of these attributes (for example: learning to predict the "glasses" attribute may be easier than recognizing "oval face") Some attributes ( Ex: "big lips", "heavy makeup") are very subjective, they are difficult for machines to distinguish, and may even sometimes confuse humans

In addition to the above problems, the training set often has the problem of unbalanced labels (for example, there are very few positive samples for the "bald" attribute), and it is very difficult to rebalance multi-label data, often balancing one attribute will affect the other. a property of balance

Images