Research On Sparse Coding Based Feature Learning Algorithms And Their Applications

Sparse coding,as an efficient and compact feature learning algorithm,has been widely used in image feature extraction.It simulates human visual s ystem by seeking a small set of atoms that can best represent input signal,which is similar to the sparse response of neurons.Thus,sparse coding can extract the intrinsic feature of images.Based on the theory of sparse representation,this thesis investigates the application of sparse model in visual feature learning,with image fingerprinting and histopathological image classification as examples.This thesis starts with a review of existing feature extraction methods,including hand-crafted and self-learning based approaches.For the latter category,our emphases are placed on sparse coding and generative model.We then analyze the biological basis of sparse coding,summarize some related theories,and provide an overview of several representative dictionary learning algorithms.Moreover,in light of the connection between sparse coding and generative model,the thesis also gives a brief introduction of generative model.Based on the above theories,we propose a hierarchical sparse coding model for robust image fingerprinting.Several measures are devised and applied on sparse coding and dictionary learning to prompt the invariance of fingerprint,such as imposing the neighborhood-priority principle on atom selection,regularizing the layouts of atoms and forcing sparse codes preserves the distance in image space.To better simulate the information processing flow of visual system,the proposed model adopts a hierarchical coding architecture.Content identification performance of the proposed work is tested on a database of 219,000 images.The error rate,which is at a level of10^-3,demonstrates the effectiveness of the proposed algorithm in invariant fingerprint extraction.Furthermore,sparse constraint is also integrated into ge nerative model,forming a sparse Product-of-Expert?PoE?based feature extraction approach.We apply the sparse generative model in histopathological image classification.In particular,two PoE models are trained to fit the distributions of healthy and inflammatory histopathological images respectively,where the sparse constraint forces the model to capture the discriminative features of histopathological images.The responses of an image under the two models are concatenated as features.Finally,patch-level and image-level classifiers are trained using the features generated by the sparse PoE models.Experimental results show that the proposed algorithm can consistently achieve accurate classification results on the histopathological images of different organs,and its accuracy surpasses a number of state-of-the-art algorithms.