Face Recognition Algorithm With Multi-information Fusion Based On Combination Of Sparsity And Collaboration

With the rapid development of information science and technology, face recognitiontechnology undertakes an increasingly important role in the field of information securitydomain, and it is also promising in security monitoring, e-commerce, security defense and soon. There are many non-ideal factors in practical applications, such as illumination, occlusion,posture, expression and background, therefore, it is important to explore a robust facerecognition algorithm. Sparse representation and collaborative representation can capture theessential features of face image, they have become the research focus due to their excellentrecognition performance. Collaborative representation needs the irrelevance of the columnvector in the dictionary. However, the face images of the different classes have the similarpattern, the relevance among the columns is high, this will increase the false classification incollaborative representation. In addition, the collaborative representation is based on the largenumber of training samples, it is hard to satisfy the simple training sample in the practicalrecognition. Aiming at problems of collaborative representation and simple training samples,we mainly focus on two aspects.(1) Compared with the sparse representation, collaborative representation can obtain highrecognition rates and low computational complexity. However, the column vector’s relevanceis high in dictionary, this will increase the false classification. Kernel function is thedimension transformation function, the feature vectors are mapped from the low dimensionalfeature space to high dimensional space, the geometry distance among the classes willincrease, which makes the linearly non-separable problems into a separable one. In order toimprove the classification precision, we propose the kernel collaborative representationclassification. The face images are mapped to a nonlinear feature space by applying KPCAalgorithm and extract the most effective features for recognition, then choose the regularizedleast squares solution for classification. Experimental results on AR database show that thealgorithm greatly improve the recognition rate under the change of occlusion, illumination,and also reduce run time than1-norm minimization.(2) Aiming at the problem of simple training samples in face recognition, the classic SIFTalgorithm can extract the stable feature descriptors which can obtain perfect matching performance when two images have the problem of translation, rotation, affine transformationand perspective, illumination changes. Face images have similar regions, therefore, the SIFTalgorithm has many false matching pairs and high complexity when it extract features andaccomplish matching. We propose a face recognition algorithm based on trace transformwhich extract local features to increase the number of training samples and increase thesparsity of the training samples. With application of one order Scharr operator and secondorder scale-adapted Laplacian of Gaussian (LoG)&Harris filter, feature points are detected.Proper functionals are chosen to perform trace transform in the circular region around eachfeature point so as to achieve feature descriptors which are invariant to rotation and scaling. Acoarse-to-fine matching strategy is conducted by applying the descriptor’s feature vectors andcoordinates. It guarantees the stability as no problem of parameter selection. The experimentalresults on ORL, AR, Extend Yale B show that this method decreases the effect caused by thevariations of pose and expression and also reduces the running time than SIFT algorithm.