Research On3D Face Detection And Recognition

The face recognition technology has great application prospects such as HomelandSecurity, Military security, Financial Security and Public Security etc. The humanbrain has natural capabilities for recognizing face easily, but automatic facerecognition is still a very difficult problem for computer. The previous reportsshowed that the research on automatic face recognition could be traced back to thelate19th Century. But limited by the constraints of various conditions at that time,there were no major breakthrough about recognition theories and applications. Until1990s, a series of excellent2D face recognition algorithms are represented, theface recognition research has became the hotpot. However,2D face recognition isinevitably influenced by the illumination, pose and expression and these factorsbecome the biggest bottlenecks for the development of the2D face recognitiontechnologies.As the3D real-time data acquisition techonology become into reality, more andmore face recognition researchers turned to the research on3D face technology. The3D face recognition technologies are expected to fundamentally solve the problemswhich troubled the2D face recognition. On account of the premise of3D face datahave been obtained, it will be less affected to the3D face recognition by the light.As the pose of3D face model has six degrees of freedom, so technology of3D facerecognition has the potential to overcome the posture changes in the technical theory.However, expression is still the bottleneck for3D face recognition. This paperfocuses on the problem of pose and expression, gives an effective solution and putsforward the key algorithm from face detection to face recognition. The maininnovative research work of this paper can be summarized as follows:1, Proposed the preprocessing methods, such as3D face detection, face cut andposture correction of3D face recognition.As the3D face data contains many of the useless data such as ears, neck,hair and clothes etc. Before face recognition, we need a preprocessing step.As the nose tip is the area which has most obvious geometric characteristicson the face surface, that it can be easily recognized. Therefore, the nose tiplocation is the first step in the face detection. By analysising the geometriccharacteristics of the nose tip region. We give three filter criteria to getthe nose tipPn oseand its normalnn ose. Secondly, locate the nose point as the center and then cut out the facial regionM faceaccording to the cuttingradius r under the geodesic distance measure. Thirdly, we get the profice C,the root of nosePb aseand the directionnb aseof the bridge of the nose. Finally,the face regoin is transformed into the unified coordinate frame.2, Proposed a3D face recognition method based on multi-mode2D images.Based on a full analysis of the3D face model, this paper attempts to reducethe dimension of3D face geometric information from3D to2D. In order to ensurethe integrity of3D geometric information, three types of2D images areintroduced in this paper. The first type is the projection depth image; thesecond type is the geodesic distance image and the last type is the mean curvatureimage. Through the above transformations, the problem of3D face recognitionis simplified as the2D face recognition.We select the2D face recognitionalgorithm based on sparse representation, and introduce the sub-image strategyin order to reduce the negative effects of face pose and expression.The finalface classification is the fusion result of all the classifications of eachsub-images.3, Proposed a3D face recognition method based on the fragmentation, weighting,partial-ICP and geodesic distance.The3D face recognition methods based on the ICP algorithm have a goodability to adapt to the face pose variations, which is a big advantage of thekind of such algorithm, but the expression variations could cause therecognition rate significantly decreased. The method proposed in this paper isto ensure the strong adaptability to the pose variations, it also can reducethe impact of expression variations for the face recognition through introducingthe following three strategies: Firstly, introduce the fragment method anddivide the face region artificially into7sub-regions, which can reduce theinfluence of local large deformation on face recognition. Then introduceimproved partial-ICP to match sub-regions. Through adjusting the p-ratecoefficient, it can greatly reduce the influence bringed by the expressionvariations. Finally, according to the geometric properties of facial surface,the sub-regions are given different weights, the purpose is to increase theweight of sub-region which has good rigidity, and reduce the weight of thesub-region which is more easily affected by expression changes. In the finalclassification, for the classification result which has poor differentiation,the geodesic distance metric is introduced as the assisted discriminatecriterion.4, Proposed a3D face recognition method based on the multi-angle projectionimages SIFT matching.Different from the general idea of the face recognition, the method doesnot demand a good orignial pose. We get other eight groups face projection imagescome according to certain rules using the original face, and form the new face images database whose size is nine times of the original one. The famousfeature matching algorithm SIFT in2D registration field has been used in thelater match procedure. Due to the SIFT algorithm has shortcomings that easilybring mismatch points, we present a method to remove the mismatch pointseffectively. Finally face classification is according to the number of matchingpoints.At the end, we give the prospect of the future work in this field.