Research And Implementation Of Face Information Analysis System Based On Mutiltask Learning

With the continuous development of artificial intelligence,the degree of social information construction is deepening,and the demand for facial attributes of human face is increasing in various fields such as education,industry,transportation,public security,etc.The application of advanced computer vision technology in various fields of facial information analysis enhances the universality and uniqueness of the application of face analysis.Face information analysis has strong biological attributes.There are mature evaluation systems and high-precision experimental results for single-task face attribute analysis algorithms.However,in practical application scenarios,it is necessary to be able to simultaneously perform multi-task facial attribute analysis techniques,but inherent visual task problems such as illumination,pose,image quality,and computational complexity make the multi-task attribute analysis accuracy dropped significantly,while the time delay of cascaded face detection and face recognition is unavoidable.Based on this,this paper firstly studies various single tasks of facial information analysis,and proposes a facial information analysis algorithm based on multi-task learning,then designs and implements a facial information analysis system based on this algorithm.The multi-task based facial information analysis algorithm and system proposed in this paper includes common facial information analysis tasks--face detection,face landmark detection,face recognition,gender recognition and expression recognition.Combined with various existing facial information analysis task algorithms,the convolutional neural network and multi-task learning method are used to fuse various facial information tasks and realize collaborative processing of multi-task facial information analysis.Firstly,in order to achieve efficient facial information analysis tasks,improve the performance of single task,and accelerate the time delay problem caused by traditional detection and recognition cascade.The face information analysis algorithm based on multi-task learning proposed in this paper uses the face information analysis task system framework with hard sharing mechanism,adopts the parallel face information analysis network structure,and uses the feature sharing method to save the network calculation amount and obtain more generalized facial information features,while implementing face detection and face recognition tasks.And the algorithm is suitable for a single constraint scenario to complete face verification in an efficient way.Secondly,in order to improve the accuracy and applicability of the facial information analysis algorithm in parallel multi-task network,other face information analysis tasks are added.Face landmark sharing is added to the face detection network branch to perform supervised verification of face detection.Gender recognition and expression recognition is added to the face recognition network with the ResNet network branch to further learn the generalization characteristics of the face region,and improve the performance of various single recognition tasks by identifying the high-low dimensional feature fusion method of the network feature layer.Finally,in order to train and evaluate the facial information analysis network based on multi-task learning,this paper has collated and tailored the public data sets of various facial information tasks,including large-scale face recognition data sets,multi-type expression recognition data sets,gender identification datasets and large-scale face detection datasets and facial keypoint detection datasets.The test of performance on the FDDB(Face Detection Datasets and Benchmark)face detection data set and the LFW(Labeled Faces in the Wild)face verification data set is carried out.The experimental results show that the multi-task based facial information analysis algorithm can achieve 98.2%accuracy in the ROC evaluation curve and 98.9%accuracy in face verification.The face information analysis algorithm based on multi-task learning proposed in this paper can effectively help the network to have better facial feature extraction ability by learning related tasks between facial attributes.Key technologies such as parallel network structure,lightweight network model,target region mapping mode,and high-low dimensional feature fusion method are used to improve the accuracy of facial information analysis algorithm,and simultaneously complete the task of multi-attribute facial information analysis.And the wide applicability of the algorithm model is verified on the designed multi-task facial information analysis system.