Discriminative and Robust Sparse Approximation for Face Recognition
Loading...
Date
item.page.authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
newline In the last few years, sparse representation methods have given satisfactory results in facerecognition. In real-world applications, face images exhibit a wide range of poses, lightingconditions, and facial expressions, causing traditional sparse-based algorithms to degrade significantlyin performance. In this work, different sparse approximation-based algorithms areproposed to improve the classification accuracy of face recognition.In this work, various issues and challenges associated with unconstrained face recognition are analyzed as well as the feasibility of using a multi-modal feature vector for face recognition. Furthermore, the efficacy of sparse-based face recognition approaches with limited training datais explored and a hybridization approach combining sparse approximation with cuckoo searchis proposed for improved performance. Using sparse approximation at multiple levels, facerecognition is implemented that accurately discriminates between individual differences and isrobust to intra-individual variation. This work also addresses the problem of handling imageswith partial occlusion by implementing sparse approximation in two distinct quadrants of a face image.In the perception of faces, both global and local features are equally important. Localdescriptors can be used to capture fine details in local variations on a face, while global descriptorscan be used to describe the structure and contour of a face. Face recognition using featurescontaining both color and texture information performs better than systems using only color or texture information. Considering this, multi-modal feature vectors are experimented with byutilizing a combination of local texture and global color features. Multi-modal feature vectorsare derived by combining very efficient local texture descriptors (Local Binary Pattern (LBP))and global color characteristics. Recognition results obtained on benchmark databases provethe value of the proposed approach of extracting multi-modal features.To improve the performan