Research On Face Recognition Algorithms From Image Sets Based On R₁-PCA

The face recognition technology is a technology to detect human face images, extract thefeatures and process the identities of recognized face by means of some computer sciencetechnologies and specific algorithms. Comparing with other bioinformatics such as iris andfingerprint, human face image has remarkable advantages including non-contact,non-mandatory, easy to perform and etc. In some areas such as image processing and patternrecognition, the research on face recognition technology has always been the front edge. Withthe upgrade of computer hardware and improvement of network communication technologies,face recognition technologies have gone through considerable development. Nowadays, facerecognition technology has been applied widely to security area such as identity recognitionand smart monitor and to media entertainment area and human-computer interaction areaincluding many kinds of smart devices, movies, games and etc.The traditional face recognition refers to face recognition from single-shot. Generally, ittakes singe-shot image as a unit to be analyzed and divided into. Also, it can classifysingle-shot image and recognize the identity with classification model by learning frommultiple images. The traditional face recognition has been developed widely and well andpossess mature theory and technologies. However, the face recognition from single-shotgenerally will not have ideal performance and effect when it takes complicated environmentalelements including illumination, posture, facial expression, age and gender into consideration.The Chinese and foreign researchers have proposed a lot of algorithms and had someaccomplishment. However, most of these algorithms cannot be implemented under thespecific and limited condition in a laboratory and cannot meet the specific requirement inrealistic applications. Especially, when the face recognition from single-shot is applied tovideo surveillance, it has low efficiency of recognition because of the unsatisfactory quality ofimages. Recently, the face recognition from image sets has got more and more attention. Thistechnology takes human face image sets consist of multiple images as units to be analyzed.An image sets of the same person generally includes a large numbers of images in differentillumination, posture, facial expression and etc. The much more information from image setsmakes the efficiency of recognition more stable. The face recognition from image sets takesimage sets consists of multiple images as a whole, models the image sets and performs theclassification by calculating the similarity or distance between specific image sets withdefined image sets. Both the learning and classification take image sets as units, which is themost prominent feature different from traditional face recognition. After reading large numbers of Chinese and foreign papers concerning face recognition,this paper makes a brief review on face recognition from image sets, creatively classifies theface recognition from image sets based on the modeling of image sets. And this paperanalyzes the strengths and weaknesses of different face recognition technologies from imagesets and raises two pivotal questions to be solved: how to model the image sets and how tocalculate the similarity and distance between them. Finally, this paper proposes a frameworkfor face recognition system from image sets.This paper takes deep research on Affine Hull Based Image Set Distanceand analyzestheoretically the reason of the sensibility to outliers—the effect from outliers amplified byL2-norm. By introducing R1-norm—principal component analysis algorithm based on rotationinvariant and estimating the robustness for affine subspace, this paper proposes R1-AHISDalgorithm which is insensible to outliers.This paper applies the Viola-Jones face detecting algorithm to the extraction of humanface images from standard video library Honda/UCSD to build the image sets database. Thispaper implements the classical Viola-Jones face detecting algorithm, improves the faceimages extracted from standard video library with proper parameters, builds the image setsdatabases to verify the effectiveness.The R1-AHISD algorithm proposed in this paper models the image sets with affine hullmodel and performs the estimation for affine subspace with R1-PCA algorithm and obtains theclassification results with the nearest neighbor classifier by calculating the defined distancebetween affine hulls. This paper performs the experiments under the noise free and noisecondition, tests separately the mainstream face recognition algorithms from image sets andR1-AHISD proposed in this paper. The experimental results show that the algorithm proposedin this paper has better recognition rate and stability and displays higher robustness in thenoise experiments.