Research On Image Quality And Face Attributes Based On Deep Learning

With the advent of the digital age,images and videos have exploded and spread.People are more likely to be attracted by images and videos than text,voice and other information.Therefore,the field of computer vision is getting more and more attention from researchers.However,images are inevitably effected by noise during generation,compression,and reception.Image quality assessment can effectively evaluate the performance of these algorithms,and they also provide the goals for algorithm optimization.It is classified into subjective quality assessment and objective quality assessment.The objective quality assessment is classified into full reference quality assessment,partial reference quality assessment and no-reference quality assessment according to whether the original image features are needed.In practical applications,no-reference quality assessment do not need the original image.So,it is more useful.In this paper,a NR-IQA model based on data enhancement and deep learning is trained,named Deep No-Reference Image Quality Assessment(DNR-IQA).Since the image quality assessment database is relatively small,we use Gaussian kernel regression to fit the degree of distortion and the differential mean opinion scores function,increase the amount of data,and train image quality assessment model based on AlexNet network.The experimental results show that the model has high accuracy and good generalization ability.In recent years,face attribute analysis system has attracted more and more attention,such as face detection,face key point detection,face pose estimation,face recognition and so on.It is widely used in various identification,human-computer interaction and other scenes.In practical applications,the effect of the face attribute analysis system relies heavily on the quality of the face image.If the face in image is evenly illuminated,the posture is correct,the noise is low,and no occlusion,the performance of the system can achieve a high accuracy.However,in the stage of shooting,transmission and reception,images are inevitably subject to image distortion,large face pose,and even occlusion.The quality of the face is low,which greatly affects the effect of the face attribute analysis system.This paper proposes a face image quality evaluation classification model based on ResNet-50 network.Compared with traditional networks,ResNet has the advantages of lower parameters,deeper network and better performance.In this paper,the face images are classified into five categories based on face image quality,including fuzzy,partial and non-face,large pose,occlusion,high quality.By filtering clear face images,the face image quality is controlled on the data source,and the accuracy of the face attribute analysis system will be improved.Experiments show that the accuracy of the model for face image quality is relatively high.Many scenes need face attributes,such as security check,identity verification,humancomputer interaction and so on.Face detection is the basis of face attribute analysis and the key point of face recognition.It is effected by many factors,such as face pose,expression,occlusion and image quality.Facial keypoints detection uses 68 points to locate facial features and contours,provides location information for face recognition,face pose estimation,face alignment and other issues.Face pose estimation can not only improve the accuracy of face recognition,but also estimate the direction of the human eye and obtain the region of interest.Therefore,face related issues have always been an important research direction in the field of computer vision.This paper proposes a multi-face detection model based on the fully convolutional network.This model replaces the fully connected layers with the convolutional layers,which greatly reduces the parameters of the model,and improves efficiency.During the test stage,there is no need to adjust the image size.Then,we propose a face attribute analysis model based on multi-task learning model.It is proposed to detect 68 facial keypoints and estimate face pose.The experimental results show that the speed and accuracy of the model are relatively high.Finally,this paper proposes to use the face image quality evaluation model to filter the images of clear faces and improve the accuracy of 68 facial keypoints detection and face pose estimation.