Research On Face Image Quality Assessment Based On Classification-performance Criterion

With the development of computer vision technologies,face recognition is wildly used in many areas for personal identification.However,current face recognition algorithms will degrade seriously in identification accuracy when the face image is low-quality.So how to capture high-quality face images from the image-collection device to improve the performance of face recognition are focused by the researchers.Image quality assessment could provide a proof for image processing system's performance via designing a reasonable algorithm to make computer can evaluate the quality of images correctly.Current image quality assessment methods judge an image quality by feeling of human eyes,however,in the pattern recognition,the most suitable image may not be the image has the best human eyes feeling.This paper proposes a new face image quality assessment method based on the criterion of the performance of face classification.Following is the details of the method:1.Propose a new face image quality assessment framework according to face recognition performance.This framework simulates 6 kinds of face image degradations model and evaluates the influence of different kinds of degradations by the accuracy of face classification.This method could assesse the image quality more objective compared to the traditional method.This paper verifies the performance of traditional methods using a new database based on CMU PIE database and chooses the DIIVINE method which is more effective in traditional methods.2.Based on the feature evaluation of face image quality based on classification performance,in order to improve the classification performance of face image quality evaluation on the basis of the feature extraction of DIIVINE method,by adding local features for face image quality evaluation,Regression feature selection and feature weighting method,a non-reference face image quality evaluation method based on statistical characteristics is constructed.Experiments show that the proposed face image quality assessment methods in this paper increases the consistency of the face classification performance and image quality score.