Face Detecion And Analysis Based On Deep Learning

The face analysis system is of great significance in security scenarios such as video surveillance.At present,under normal lighting conditions,the visible light-based face analysis system has achieved excellent results in various tasks of the face,but the analysis efficiency needs to be further improved.In addition,in some challenging surveillance scenes,such as in a dark environment,because visible light cameras cannot collect clear facial images,visible light-based facial analysis systems are limited.In this case,only infrared cameras can collect clear facial images.Therefore,cross-modal face recognition based on infrared-visible light has an important complementary role and strategic significance for monitoring security scenes.In response to the above problems,this paper first proposes a multi-task heterogeneous cascaded convolutional neural network based on visible light to improve the analysis efficiency of the face analysis system in the surveillance scene.Secondly,for low-light or dark environments with limited illumination,this paper proposes a globallocal constrained face generation confrontation network GCF-GAN for cross-modal face recognition,which effectively improves the generation results from near-infrared face to visible light,and improves the accuracy of cross-modal face recognition.In a surveillance scene with normal lighting,a visible light-based face analysis system not only needs to achieve high precision,but also be as efficient as possible.In order to improve the analysis efficiency of the face analysis system in the surveillance scene,this paper proposes a universal framework,termed as Multi-task Hybrid Convolutional Neural Network(MHCNN),for joint face detection,facial landmark detection,facial quality,and facial attributes analysis.MHCNN consists of a high-accuracy single stage detector(SSD)and an efficient tiny CNN(T-CNN)for joint face detection refinement,alignment and attributes analysis.Though the SSD face detectors achieve promising results,we find that applying a tiny CNN on detections further boosts the detected face scores.By multi-task training,our T-CNN aims to provide five facial landmarks,facial quality scores,and facial attributes like wearing sunglasses and wearing masks.Since there is no public facial quality data and facial attributes data as we need,we contribute two datasets,namely Face Q and Face A,which are collected from the internet.Experiments show that our MHCNN can significantly reduce the false detection of single-stage face detectors under high recall rates on FDDB,and gets reasonable results on Face Q and Face A.Although face recognition technology based on visible light has developed rapidly and has abundant practical applications,in low-light or dark environments with limited light,due to the high cost of infrared face collection and the scarcity of data,cross-modal face recognition based on infrared-visible light faces great challenges.In response to this problem,this paper proposes a global-locally constrained facial generation Adversarial network GCF-GAN for cross-modal face recognition,which effectively improves the generation results from near-infrared face to visible light,and ineffectively improves the accuracy of cross-modal face recognition on the CASIA NIR-VIS 2.0 dataset.Since the discriminability of human faces is mainly reflected in the differences in core areas such as eyes,nose,and mouth,GCF-GAN strengthens the local area by imposing additional local constraints on the core areas of the facial features in the process of generating faces.The details are generated to generate a more realistic visible light face,which improves the accuracy of cross-modal face recognition.In addition,because there is no large enough thermal infrared-visible face dataset on the market.This paper contributes a large-scale,multi-attribute,multi-posture polarization imaging thermal infrared-visible face dataset SIAT IR-VIS 1.0,which is a data set with great practical application value and strategic research significance.