Study On Algorithms For Image Classification Based On Sparse Coding

With the rapid development of the Internet, there is an explosive increasing number of images on the web, which causes the problem of how to use computer technology to manage these images effectively a attractive research issue. The main work of image classification is to study how to extract effective features of the images, then classify and recognize them automatically. Moreover, image classification accuracy has been greatly improved, after applying the sparse coding method in the field of image processing.Sparse coding is an effective feature selection method. It firstly trains extracted image features to get a overcomplete dictionary. And then obtains a dictionary-based reconstruction coefficient by encoding. Finally, obtain the feature vector representation through the spatial pyramid model. The key point of this method is making use of the overcomplete dictionary's linear combination of basis vectors to represent the features of images, and each feature coding contain only a few non-zero elements which reflect its sparsity. Compared with the traditional method which generates overcomplete dictionary by K-means, sparse coding works better because it weights each basis vector which represents a feature while dictionary's learning and optimizing. But there still exist some shortcomings that sparse coding can hardly overcome. Sparse coding ignores the correlation between various features while encoding, which results in similar features getting a completely different coding and inaccuracy when classifying images.According to above discussion, the main work and results in this thesis are as follows:First of all, this thesis introduces the basic framework of image classification, in which we will discuss the basic ideas and specific processes of several common feature extraction methods, feature representation methods and linear classifiers, then compare their advantages and disadvantages. Finally, we will verify their accuracy in image classification through a large number of experiments.Next, this thesis proposes an improved sparse coding algorithm in this paper--lifetime sparsity constraint sparse coding method(LTCSC), which adds the basis vectors' activation constraint in the original object function of sparse coding. The new object function restricts the activation times of each basis vector, selects basis vectors representing images evenly to effectively guarantee the lifetime sparsity of the coding.The LTCSC method improves the accuracy of image classification effectively.Finally, this thesis will also propose an improved algorithm of dictionary learning in this paper-- clustering based optimize dictionary learning method(CBODL).Based on the traditional K-means clustering, CBODL optimizes the method of choosing the initial cluster centers and chooses vectors which are less relevant to be the next new cluster centers. By this way, both the quality of the dictionary and the precision of image feature quantization are improved, thus effectively improving the accuracy of image classification.The two algorithms proposed in this thesis have been proved to be able to get good results in image classification through a lot of experiments on the datasets Caltech 101 and Scenes15.