Study Of Automatic Eye Location Method For Face Recognition

In the research of machine learning, face recognition is a hot research area. Thegeneral model of the face recognition system is divided into three steps: face detection,feature location and extraction and recognition classification. The third step featurelocation and extraction is the basis of the entire face recognition algorithm, and itsaccuracy has great impact on the performance of the algorithm. So it is of greatsignificance. But usually in the process to extract feature points, it can’t obtain verysatisfactory results because of the complex factors of the face, such as the illuminationchanges, cover glasses or hair, facial expressions and other issues. In order to locatethe facial feature points, the most critical matter is to predict the location of the eyesor mouth, eye location has also become a hot research topic.The main research of this article is on the active shape model. First we design andachieve four basic face recognition algorithms: principal component analysisalgorithm, linear discriminate analysis, the image difference classifier, elastic bunchgraph matching algorithm, followed by the improved ASM algorithm with featurelocation calculation process. ASM multi-template matching method is proposed tolocate the position of the human face feature points more precisely, and extracted theprecise human eye coordinate template based on positioning facial feature points.Finally, we operate the multi-angle experiments with manual calibration of the humaneye coordinates and automatic extraction of the human eye coordinate. The study willinclude the following three aspects:1) On the basis of the standard active shape model, in order to avoid the defect thatthe ASM algorithm is easily falling into local minima, the ASM algorithm is improvedto multi-template matching method. Depending on the different states of the eyes andmouth, we split the multi-template matching and shield useless information in thesearch process by adding texture information and other useful information. Theasymmetric sampling strategy is adopted on the feature points of the external edge ofthe face contour in order to improve the performance of the ASM algorithm and thepositioning accuracy of facial feature points. The human eye coordinate automaticallyextracted by ASM algorithm is much precise than conventional eye locationalgorithm.2) Improve the traditional active contour models by re-division of the external force of the active contour models into the balloon force and image force. Let the improvedBalloon model apply to the human eye location to obtain better performance. Withgood pretreatment conditions, the Balloon model calculation compute fast, only needto calculate the balloon force of few important control point and its adjacent areapoints. What’s more, the Balloon model itself does not converge with thecharacteristics of the furrow, and it can optimize the silhouette closer to the humaneye and avoid the initial contour sensitive problem.3) Reference CSU face evaluation system, summary of commonly used facerecognition algorithm: principal component analysis algorithms, linear discriminantanalysis, image difference Bayesian classifier and elastic bunch graph matchingalgorithm. We improve the Mahalinobis L1Distance, Mahalinobis L2distance, aMahalinobis cosine distance, and the Young’s angle distance based on theMahalanobis (Mahalinobis) distance metric, and then does the analysis combiningwith the algorithms and metrics. In order to determine whether the automaticrecognition of the human eye coordinate can achieve the same effect of manualcalibration of the human eye coordinates, in addition to judge the eye average pixelcoordinates between the automatic recognition and manual calibration, multi-angleevaluation is proposed. According to the CSU evaluation system, the eye coordinatesare compared between automatically recognize and manual calibration recognize witha variety of different recognition algorithm and the type of distance measures. As longas the correct rate is within an acceptable range, we believe that the automaticrecognition of the human eye coordinate has a certain effect. In some certaincircumstances, automatic recognition can replace the manual calibration.