Font Size: a A A

Face Detection And Recognition Approaches Based On Sparse Representation And Dictionary Learning

Posted on:2019-08-27Degree:DoctorType:Dissertation
Country:ChinaCandidate:J J LiuFull Text:PDF
GTID:1368330572968878Subject:Control theory and control engineering
Abstract/Summary:PDF Full Text Request
Face detection and recognition techniques have been widely employed in many fields,such as identity checking,human-computer interaction control,and video surveillance,etc.However,due to the complicated situations of changing illumination,angle,pose and expression,as well as high-dimensional data and limited training samples,the current face detection and recognition techniques are still confronted with many difficulties and challenges.In the past few decades,a large number of face detection and recognition algorithms based on image processing and machine learning have been proposed and successfully applied,in which sparse representation algorithm has become a popular topic because of its excellent performance of data dimensionality reduction,feature extraction and classification.Through the analysis of the development of face detection and recognition methods at domestic and foreign research status,this thesis promotes the theoretical and experimental research on face detection and recognition methods based on sparse representation and dictionary learning,which has expansive prospect in face recognition.An improved face detection method based on region similarity is proposed to improve the low efficiency of face detection in video sequence images.Firstly,the key information between frames is extracted from the inter-frames based on the contour and position of face,and then the suspected face zone and non-face zone are divided by the face search window.Secondly,this method selects and removes regions based on the results of region similarity from texture and color,and the non-maximum suppression algorithm is utilized to mark the face detection window.Compared with the traditional exhaustive search strategy,this method significantly reduces the selective regions,fast removes the irrelevant region from one frame and greatly improves the detection efficiency.To improve the low efficiency of the high-dimensional face image processing and recognition,a classification method based on manifold constraint dictionary learning is proposed.Firstly,this method utilizes nonlinear dimensionality reduction method to project each open interval to a low dimensional linear space,employs the adjacent points to approximate the projection points of the test samples,and exploits the neighborhood constrained conditions to construct a non-convex objective function.Its analytical solutions can be obtained by the Lagrange multiplier method,and alternating direction method of multipliers is used to optimize the objective function,which can reduce the computation time and the storage space.Secondly,by constructing a structured dictionary related to the category,the sparse representation coefficients and reconstruction errors of the testing sample can be used for classification.Experimental results show that the manifold constrained conditions can enforce the intra-class distance,improve the discriminative ability of the structured dictionary and performance of classification.In accordance with the shortcomings of traditional dictionary learning such as K-SVD which cannot be discriminated effectively under small sample datasets,a face recognition method based on improved dictionary learning and joint sparse representation classification is proposed.Firstly,the embedded algorithm of OMP in K-SVD is replaced by an improved iterative hard threshold algorithm,and an improved dictionary learning algorithm(IIHT-KSVD)is obtained by introducing discriminant constraint condition.The robust principal component analysis method is employed to extract the inter-class and intra-class variation face information in the test samples,and the improved algorithm is used for dictionary learning to obtain a compensation dictionary with better expression ability for face variation.Then,the improved joint sparse representation(IJSR)classification is obtained by embedding the IIHT-KSVD approach.In classification,the compensation dictionary and training samples are sparsely employed to reconstruct the test samples,which compensates the lack of changes in the training samples.The experimental results show that the proposed method achieves better recognition results under a less training samples and complicated illumination conditions.To solve the problem of face recognition in multi-angle datasets,a discriminable multi-feature and joint sparse representation method is proposed based on local and global features.Firstly,the local features(HOG)and global information of the face image are obtained to construct the fusion feature,and the sparse dictionary is optimized by introducing the discriminant loss function.Then,by using the L2,p mixed norm,a joint sparse regularization term is proposed to represent and classify the local and global features.The experimental results show that this method can reduce the dimension of fusion features,excavate the correlation between various features,and make the learning sparse coefficient more discriminative to adapt the requirements of multi-angle face recognition.An improved method of correlation filter is proposed to improve the low accuracy of face feature extraction and robustness of recognition algorithm in multiple datasets.Firstly,a convolution sparse coding features with manifold constraint is employed to extract the details of face feature,and the Laplace feature graph is introduced to restrain the feature,which maintains the similarity of neighbors,preserves the original spatial information,and enhances the description of the feature description.Then,adopting the adaptive multiclass correlation filter,the correlation between features is captured and recognized.By constraining the filter template of the strongest response,the maximum interval and the manifold subspace constraints are increased to promote the fast convergence of ADMM optimization.Experimental results show that this method can effectively solve the problem of multi-feature classification and under-fitting,and improve the accuracy and robustness of face recognition in multi-samples.
Keywords/Search Tags:Face recognition, Sparse representation, Dictionary learning, K-SVD algorithm, Regional similarity detection, Manifold dimensional reduction, Correlation filter, Convolution sparse coding
PDF Full Text Request
Related items