Hand-written Digit Recognition Based On Kernel Methods And Nonlinear Sparse Representation

Handwritten Digit Recognition is the core technology in processing some datainformation, such as Automatic Examination Paper Marking, statistical reports and otherdata entry systems, therefore, Handwritten Digit Recognition is of great significance.Sparse representation and dictionary learning has been widely used in image processing,pattern recognition and many other fields. However, traditional sparse representation anddictionary learning theory are based on a linear model, thus almost always inadequate forrepresenting non-linear structures of the data which arise in many practical applications. Inresponse to this shortage, considering the sparse representation and dictionary learning ina high-dimensional feature space and thus patterns will be separate linearly. This paperconsiders the nonlinear sparse representation used in the field of Handwritten DigitRecognition, mainly to complete the following tasks:First, based on the existing nonlinear dictionary learning methods, this paper studiesa method called the Meta-KKSVD to update the dictionary atoms, kernel dictionarylearning algorithm. Experiments show that, compared to traditional dictionary learningalgorithms and some existing nonlinear dictionary learning algorithm, the Meta-KKSVDkernel dictionary learning method can provide a better performance in handwritten digitrecognition.Secondly, the SRC fails to consider the distinctiveness and similarities of differentcharacteristics in the coding and classification stages and this paper presents a nonlinearsparse representation, which is called kernel-based relaxed collaborative representationalgorithm, and achieve nonlinear sparse representation in a high-dimensional feature space.Compared to RCR and other kernel-based methods, this algorithm effectively improvesthe recognition accuracy of the benchmark handwritten digits datasets of Mnist and USPS.Finally, taking into account of the rank minimization problem to nonlinear patternrecognition, here researches a nonlinear kernel dictionary learning method combining withlow-rank decomposition. The algorithm applies the low-rank images to the nonlinearkernel dictionary learning process, and ultimately gets a more discriminative dictionary. Compared to the original KKSVD method, the low-rank based KKSVD learning methodslead to a much higher classification accuracy in the dataset of USPS.