Research Of Face Detection&Recognition And Its Application Under Complex Enviroment

As an inherent attribution of human, biometrics (including behavioral characteristicssuch as voice, signature, gait, etc, and physiological characteristics such as fingerprint,palmprint, face, iris, etc) with strong stability and individual differences, are ideal bases forautomatic identification. As an important biological characteristic, face has advantages inmany aspects such as universality, uniqueness and acquisition. Face identification, as one ofthe most successful branches of personal identification, has great academic research value andmarket application prospect.This thesis mainly contributes on face detection and recognition problems. Intwo-dimensional (2-d) face identification, there are some inherent defects under complexenvironment. In order to effectively deal with this challenge, this thesis focuses on faceidentification by fusing2-d and three-dimensional (3-d) face data. In the algorithm design, theeffectiveness for existing public face database and practical requirement are fullyconsidering.The work and the innovation in this thesis can be summarized as following:1. Civil-level3-d data acquisition devices, taking kinect as example, subject to theirpoor performance of civil-level acquisition device and complex environment, which isinevitable for acquisition data with noise, holes and so on. For2-d image data, the commonde-noising, filtering algorithms can improve the data quality. The presence of holes iscommon in3-d image data. For small holes, the regional growth and smoothing algorithm canfill the holes. Meanwhile multi-frame fusion3-d data based on local search strategy is appliedto solve the larger holes. Experiments show that the proposed method for face data prepropesscould deal with the above mentioned problems.2. Face detection is the premise and basis of face recognition and related applications.Most previous studies focused on the constrained conditions in face detection. Aiming at thecomplex conditions of face detection, considering the inherent defects for algorithm based on2-d face image, this thesis proposes a method by fusing2-d and3-d data to solve facedetection, under complex conditions such as weak light, partial facial features and multi-pose.Under normal light condition, the Dynamic Time Warping (DTW) algorithm is applied toface contour extraction. At the same time, The real face is filtered based on the variance offace depth information.In order to solve the problem of human face detection under weak light, a quick locatingmethod based on thedepth information by Kinect device is proposed. Firstly,2D Chamfer Match method is used to rapidly preliminary locatethe human face. Secondly, region-growingalgorithm is applied on the preliminary locating face. Finally, combined withextended Harrfeatures, the Adaboost algorithm is used to train the strong classifier under weak light, whichrealizes theprecise face detection. Rationally utilizing the depth information to detect thehuman face would reduce the scope of thesearch window, which accelerates the detectionspeed. The experiments show that the proposed method not only solves theproblem of facedetection under weak light, but also is suitable for the face dection under the normal light,which has strongrobustness and timeliness.The problem of face detection based on partial feature is divided into two cases: undernormal light condition and under weak light condition. Under normal light, Firstly, the facialfeature extraction based on HSV color space has been studied. Then, the task of detection hasaccomplished by combining feature points’ space curvature with relationships. In the weaklight condition, firstly, based on sufficiently study on the nose weight in facial features, theeffective energy concentrated nose candidate regions are found by integral projectionalgorithm in3-d data space. Secondly, the nose tip is located by curvature constraints. Thirdly,the other facial features surrounding the nose tip are extracted by the improved search strategy.Finally, the purpose of detection is achieved by the cascade classifier.For multi-pose problem, Firstly, the training data is automatic classified by fuzzy c-meanscluster algorithm. Then, the independent classifiers trained by float boost learning inseparated sample subspace. Finally the multi-pose face detection purpose isrealized.3. As for face recognition, this thesis focuses on the method based on information fusion.Under normal light condition, in order to break through the constraint of2D face recognition,3-d face recognition based on Kinect instrument with information fusion is proposed. Animproved Local Binary Patterns (LBP) algorithm is used to face recognition, where theweight value distribution is determined by analyzing the color image information and facephysiological characteristics. Experiments show that the proposed method for face detectionand recognition has the use value and has improvement in real-time capability and robustness.For the problem of partial facial feature, the systematic study of face recognition is performedby combining nose and other facial features based on3-d LBP algorithm.4. Facial features are complex data sets, including personality traits and common traits.Identification of individuals based on facial features is the study of the personality problem,and gender recognition is the study of common problem. According to the differences of thephysical characteristics for face different gender, the face different gender distribution pointsare measured, including the points’ gradient features and distance features. Combination of two types of features, the model with higher recognition rate has been trained based onrandom forest algorithm for gender recognition. In addition, the impact of features missing foralgorithm is discussed, including features missed in rigid region and flexible region.