The Research Of Face Detection And Face Recognition Under Complex Background

Face recognition problems have become to be a hotspot of computer vision andpattern recognition area recently. And face recognition is the most convenient,intuitive, and acceptable authentication technology. Therefore, face recognitiontechnology has a high research value. At present, there are a lot of algorithms forface recognition and these algorithms perform very well. Facial feature will changelargely with different light, different times, different posture and differentexpression and will be affected by complex background, glasses, beard and so on.The difficulties caused by these problems for face recognition are a huge challengeto the researchers. The face recognition system is not able to recognize faces anddistinguish identities under any natural conditions and any gestures. So facerecognition algorithm with a higher recognition rate, a faster identification and abetter robustness is our goal.Face detection is an essential first part of a face recognition system, and has adecisive impact to face recognition. Skin color and hair are important features forhuman and they are not affected by facial expressions and gestures. In addition, skincolor and hair are clustering in chrominance information for people of different races.So skin color and hair are suitable for face detection. In this paper, we propose a facedetection method based on skin color and hair using HSB color space and we callthis method the SCHFD. A lot of face detection experiments show that the SCHFDmethod not only retains the correct accuracy of the algorithm based on only skincolor but also excludes more non-face area and reduces the error accuracy of facedetection.In this paper, we study the existing face recognition algorithms and propose anovel two-step sparse representation method. We call this method the TSSR. Thefirst step of the TSSR performs coarse classification and the second step of the TSSRmakes ultimate classification. The first step first seeks to represent the test sample asa linear combination of all the training samples and chooses a combination of c₁classes that has minimum error with the test sample from all training samples. Andwe assume that the test sample is from one class of the c₁classes. The second steprepresents the test sample as a linear combination of the determined training samplesthat are from c₁classes and uses the representation result to perform classification. As the TSSR uses the training samples that are most similar to the test sample torepresent the test sample and can eliminate the interference of the training samplesthat are far away from the test sample. We could obtain the higher classificationaccuracy by excluding the contribution made by the training samples which are faraway from the test sample. A number of face recognition experiments show that theTSSR performs very well.