Research On Facial Attribute Recognition Algorithms Based On Deep Multi-task Learning

Facial attribute recognition is a hot and challenging research topic in the field of computer vision,which has attracted the attention of many researchers.Facial attributes usually represent a detailed description of a person’s identity,such as the person’s age,gender,facial shape,etc.Therefore,facial attributes are widely used in face retrieval,face recognition,video surveillance and image generation,which are viewed as mediums to describe important facial information that human can understand.However,in practical applications,facial attribute recognition is still affected by face occlusion,illumination change,low resolution and other problems,thus leading to the decline of recognition accuracy.Recently,deep learning has achieved good application prospects in various pattern recognition and computer vision tasks,including facial attribute recognition.Because facial attribute recognition needs to predict multiple different attributes at the same time,and there are the correlations and differences between different attributes.Therefore,multi-task learning has been applied in the facial attribute recognition task to improve the accuracy of facial attribute recognition.It is a research work of both theoretical significance and practical value to use deep multi-task learning network for facial attribute recognition.The main research contents of this thesis are as follows:(1)In this thesis,a deep dual-path network(DPN)is proposed for multi-task learning based facial attribute recognition.In the multi-task learning framework,the method is proposed to fully consider the learning differences of different attributes,so as to divide facial attributes into partial attribute groups and general attribute groups.Two different sub-networks are designed according to different learning characteristics so that the two groups of attributes can extract features with different input scales and network depths.In addition,this thesis proposes an adaptive loss penalty strategy to supervise the learning of the model during training,which makes the model pays more attention to the training of difficult samples and at the same time alleviates the impact of the class imbalance in the dataset for attribute recognition performance.The DPN method achieves good attribute recognition performance in challenging mainstream facial attribute dataset,which verifies the effectiveness of the method.(2)This thesis further proposes an attention-aware parallel sharing(APS)network for multi-task learning based facial attribute recognition.The traditional multi-task learning based facial attribute recognition networks are mainly a serial sharing structure.However,their optimal branch nodes are difficult to select for multiple tasks,the deep features of the network are not fully shared,and only the network features at the end of each branch can participate in the final attribute recognition.This thesis proposes an attention-aware parallel sharing network to improve the serial sharing structure.In addition,an attention mechanism including multi-feature soft alignment modules is proposed to highlight the region of interest of the model,taking into account the global and local features of different network levels,as well as the relationship between the shared network and the task-specific network.Finally,the APS method achieves better recognition accuracy than several state-of-the-art facial attribute recognition methods on the challenging facial attribute public datasets.