Face Image Enhancement And Quality Evaluation For Recognition

As the most significant characteristic of an individual,face recognition is of great significance in identity recognition,and it is an important identity authentication method nowadays.In recent years,face recognition technology has had a huge impact on our daily lives.Especially in security monitoring,face recognition technology is an important safeguard to protect people's safety.However,in practical applications,due to the diverse and changeable environments of the face recognition system,and the influence of the transmission method,environmental conditions,equipment configuration,and human factors,the collected face images often cause uneven illumination,more noise,low definition,and incomplete faces,etc.,this has greatly affected the follow-up face recognition process.Therefore,in the field of surveillance,solutions are proposed to change or improve the quality of input images,which can effectively improve the accuracy of face recognition and identity authentication.In the monitoring system,the accuracy of the face recognition algorithm is mainly affected by three aspects: low or uneven brightness for the collected images,low sharpness,and incorrect face posture.And because of the uncontrollable natural lighting in the actual scene,it is more common for the collected images to have lower brightness or uneven brightness.In images with low brightness,human faces usually lose a lot of information,and the features are not obvious,which is not conducive to identity authentication and recognition.Therefore,this paper research face image quality evaluation and face low-illuminance image enhancement which proposes new solutions for improving the accuracy of face recognition algorithms under surveillance systems.(1)A face image quality evaluation method based on TOPSIS(Technique for Order Preference by Similarity to an Ideal Solution)algorithm.To filter the face images collected by monitoring and select the most suitable image for face recognition,this paper conducts a comprehensive evaluation from the three aspects of face pose,image brightness,and image clarity.First,Base on the relationship between facial texture and ratio and yaw angle,evaluate from the three dimensions of pitch angle,yaw angle,and roll angle.Then evaluate the image clarity using Tenengrad algorithm.Evaluate the brightness by converting RGB images into YIQ color space.Finally,use the TOPSIS algorithm for an overall comprehensive evaluation,and select from them the best image for face recognition.Experiments have proved that the results obtained by this method are consistent with the subjective evaluation results,and the face image quality evaluation based on the TOPSIS algorithm not only considers the face posture,image clarity and brightness,but also it can set the corresponding weight ratio according to different scenes and select the best image based on the actual environment and the collected images.This paper uses the TOPSIS algorithm to study the image quality evaluation problem for the first time.This is also a meaningful attempt in the face image quality evaluation,and it can provide new ideas for the face image quality evaluation problem in the future.(2)Aiming at the problems of uneven brightness,more noise,and lower resolution of the monitored images,this paper proposes a low-illuminance enhancement algorithm and a superresolution reconstruction algorithm for face images.Aiming at the face images collected under the conditions of weak illumination or uneven illumination in real scenes,this paper proposes a low-illuminance image enhancement model,which can achieve end-to-end training.The model generator network refers to the U-Net network,adds jump connections between symmetrical network layers,and extracts the brightness component as an attention map,which is added to the jump connection.The upsampling layer in the U-shaped network uses deep separable convolution.The model uses a global-local dual discriminator structure,a supervised generation network,and the two play games with each other,and finally,achieve the enhancement of the low-illumination image of the face.The network does not require paired low-illuminance data sets for training,which solves the dependence of traditional deep learning networks on paired data sets,and the size of the model is smaller than the general model,which can be better deployed to the end side and applied to actual scenarios in.Experimental results show that the model can enhance low-illuminance images well,avoid local overexposure or underexposure,and run faster than the traditional model.The super-resolution reconstruction of face images mainly refers to the design of Dense Net and Res Net,combined with the generation of confrontation networks,added a noise reduction module,and used multiple fusion loss functions to construct a face image super-resolution reconstruction network.The network can not only reduce the noise of the input image but also perform super-resolution reconstruction of the lowresolution face image in the surveillance system.The experimental results show that the face image reconstructed by this algorithm is closer to the original reference image,and the image details and textures are clearer than the traditional algorithm.