Hierarchical Sparse Representation For Feature Extraction And Its Applications

Feature extraction algorithms considered as a pillar key task to make vision model-ing systems fully operative. It has been effectively utilized to minimize the computation difficulty and to perform ideal classification through extraction of the significant pattern in-formation. An essential component of a successful classification system is the selection of an effective object features. Throughout this thesis, we seek to tease out basic principles that underlie the recent in hierarchical feature extraction method.The first part of this thesis presents a novel method for feature extraction by a neural response happens to be introduced by merging the hierarchical architectures with the sparse coding approach. Regarding the proposed layered model, at each layer of hierarchy, sparse coding and pooling operations have been utilized. Although the sparse coding is employed to resolve extremely difficult sparse feature representations, the pooling process by com-paring sparse outputs is utilized to determine the match between stored prototypes along with the input sub-image. It is suggested that the importance of the ideal matching has to be maintained and also discarding the others templates. The proposed model is applied and confirmed in image and speech recognition (on isolated word vocabulary). Experimental results yielded with different specifications show that the proposed method contributes to obtain more effective features compared to other methods.The Second Part of this thesis presents a new method employ formulated Hierarchical Sparse Method (HSM) to increase recognition rate of offline handwritten Arabic character. The proposed method is significantly fast and accurate. The formulated concept of this al-gorithm is generally specified as scaling and k-mean techniques; for scaling level, the key benefit could be to withdraw the attributes in higher numeric ranges dominating those in s-maller numeric scales. Concurrently, k-mean is employed to get an effective template selec-tion approach used in the derived kernel for improving the object recognition performance. The results of experiments consist of comparing the developed algorithms along with other present methods are outlined. Experimental results show that our method is simple and its localization accuracy is high compared with state-of-the-art algorithms.The third part introduces a new algorithm for template selection based on the entropy concept. Algorithm suggests picking of the template of more information and discards the templates of less information. The proposed method provides HSM with better discriminative ability. Experimental results show that the introduced method achieves good performance in template selection with fewer computation processes and shorter time.