Research Based On SIFT Feature Extraction In Face Recognition Algorithm

Face recognition is a hot issue in the fields of pattern recognition, computervision and other areas, and good prospects of face recognition technology have beenattracting more and more attention from researchers. The general process of facerecognition includes face image preprocessing, face detection, facial feature extractionand face image matching and recognition. Face recognition is a biometric recognitiontechnology by extracting information on human facial features to conduct identityrecognition. Compared with iris recognition, fingerprint recognition and DNA andother biometric technology, face recognition is a friendly interactive biometrictechnology which is natural, convenient, non-contact, non-infringement, and can noteasily be perceived. Face recognition technology is widely used in informationsecurity, public security monitoring, finance, enterprises and access control systems,and smart cameras, and many other fields.Although face recognition technology has made significant achievements andtechnological breakthroughs, it is still a more difficult task. Face images are oftenaffected by the impact of complexity of the environment, such as noise, illumination,posture, facial expressions, partial occlusion, rotation, scale changes, and a certaindegree of affine transformation. When suffering from the impact of complexity of theenvironment, the algorithm of face recognition based on geometric features, algorithmof face recognition based on templates, and face recognition algorithm based onsubspace such as PCA, LPP, NPE, etc., will be a sharp decline in recognition accuracy.The method based on local image feature points has good adaptability to thetransformation of the complex background, and it is regarded as the mainstream anddevelopment trend for face recongnition.The scale-invariant feature transform (SIFT)proposed by DAVID G. LOWE is a local feature points extraction alogorithm, whichachieves quite well results in the field of image recognition. SIFT feature extraction isintroduced into the face recognition in this article, and some improvements are made on the original SIFT algorithm, so face recognition algorithm based on SIFT featureextraction is proposed.The main work and innovation of this paper is as follows,1. Propose the gradient domain image enhancement combing visual perceptioncharacteristics, SIFT feature extraction algorithm. The gradient domain imageenhancement method which possess visual perception characteristics conducts humanface image processing, highlight the important information of face images, and atthe same time inhibit the information which are not interested. This method willdetect the local gradient saliency of face image, and enhance the human face on thegradient domain, and improve the accuracy of extracting SIFT keypoints.2. Introduce a kind of fast EMD distance to replace the original Euclideandistance to detect the dissimilarity between keypoints, and improve the matchingstrategy of the SIFT algorithm. SIFT description vector of each key can be seen as aprobability distribution, and it’s more accurate and consistent with the characteristicsof human perception to use the EMD distance to detect dissimilar between the twokeypoints. Because of EMD time complexity is relatively high, the fast EMD distanceis adopted to detect the dissimilarity between any two keypoints. The improvedalgorithm can effectively avoid the problem of inaccuracy faced with quantization,deformation and occlusion.3. In this paper, a lot of experiments are conducted on the AR face database andgive effective verification for the proposed improved algorithm. Made a lot ofexperiments to verify the proposed improved algorithm robustness against noise,illumination, scale changes, rotation, and a certain degeree affine transformation.In this paper, the comparative experiments between the original SIFT algorithmand improved SIFT algorithm are conducted on the AR face database. The resultshave demonstrated that the improved algorithm possesses good robustness in accuracy,keypoints detection precision and impact of complex enviroments. Experimentalresults show that the improved algorithm proposed in this paper is effective..