Research On Multi-Label Face Attribute Recognition Algorithm Based On Deep Learning

Face attribute recognition technology is a research direction in the field of computer vision,aiming to automatically analyze various facial attributes of a face image through computer analysis,such as gender,age,race,expression,and accessories,etc.Face attribute recognition has a wide range of applications in real life.This article mainly focuses on the current research status of face attribute recognition technology and carries out research and design on a multi-label face attribute recognition algorithm based on deep learning.The improvement is made from aspects such as attribute correlation mining,attribute grouping,attention mechanism,adaptive loss weight,and model lightweight to optimize the performance of the face attribute recognition algorithm,aiming to improve its comprehensive performance.The main contributions of this article are as follows:Firstly,to solve the problem of insufficient attention and utilization of attribute correlation in face attribute recognition algorithms and the problem of low recognition accuracy caused by uneven distribution of attribute samples in the dataset,a method combining attribute labels and mutual conditional correlation is proposed to achieve face attribute recognition.A new loss expression is constructed by using the mutual conditional probability between each pair of attributes to supervise the model training process,which encourages the model to focus more on the correlation between attributes during the learning process and better ignore the impact of sample bias and data imbalance.Secondly,in order to better enhance the network’s focus on specific attributes,a new attention method called Three Pooling CBAM(TP-CBAM)is introduced in the network model to strengthen the model’s focus on the target area.In addition,multi-task learning is carried out using a grouping learning approach.Attributes are grouped according to label correlation,and each group corresponds to a task.The network is set as a multi-branch structure for multi-task learning.Finally,considering the impact of task weight on model performance,a dynamic exponentiated weighted method is proposed to balance the loss between different tasks and automatically adjust the relative weight of tasks to optimize the model.The proposed method can effectively improve the network’s focus on the target area and achieve good performance in multi-task learning.Finally,considering that the current face multi-attribute recognition models are relatively large and not suitable for application on small mobile devices,a knowledge distillation method for compressing the model is proposed in this article.This method uses Res Net50 as the teacher model and an improved dp-Res Net18 as the student model.The response-based soft target distillation method is used to transfer information between the teacher and student.At the same time,by introducing depth-separable convolution into the Res Net18 network,the model’s parameter size is further reduced.The experimental results show that the proposed method can significantly reduce the model size while ensuring accuracy and achieve model lightweight,which is convenient for deployment on small devices.