Neural Response And Sparse Representation Based Algorithms For Image Classification

Image classification is a hot topic in the field of machine learning and computer vision and has been widely used in all kinds of real-world applications, including security of na-tional defence, industry informatization, medical engineering and internet technology. As images tend to be affected by various factors, such as light conditions, view points, rotations and scalings etc., image classification is actually very challenging. Therefore, developing efficient and robust image classification algorithms is meaningful in terms of both theory and practical applications.Extracting representative features from images is the key part of the image classifica-tion process. To obtain effective image features and solve the problems caused by lacking of labeled training data, this thesis focuses on investigating robust and discriminative algo-rithms for image classification. To this end, we propose several algorithms based on the neural response model and sparse representation method. Experimental results demonstrate that the algorithms can achieve excellent performance for image classification tasks. The content of the thesis is as follows:Firstly, we propose a sparse-based neural response algorithm for image classification. The algorithm is a multi-layer structure which tries to simulate the hierarchical structure of the visual process in human visual cortex. It is recursively constructed by alternating between non-negative sparse representation and maximum pooling operation. The non-negative sparse representation is utilized to extract semantic features from images, while the maximum pooling operation makes the algorithm invariant to translations. To further en-hance the performance of the algorithm, we propose two effective template selection meth-ods by taking the structure of the algorithm into consideration. We conduct experiments on public datasets, and the results show that the proposed algorithm can improve the per-formance of the original neural response model largely and perform excellently in image classification tasks.Secondly, we propose an algorithm with multi-layer structure based on the neural re-sponse model by introducing the theory of ELM (extreme learning machine). The algorithm includes two stages, i.e., the multi-layer ELM feature mapping stage and the ELM learning stage. Particularly, the first stage is a multi-layer structure which is constructed by alternat-ing between feature map construction and maximum pooling operation. It should be noted that randomly generated weights are employed to construct the feature maps and no need to be tuned during the learning process. Consequently, the proposed algorithm is highly simple and efficient regarding the computational complexity. During the ELM learning stage, we propose to utilize the elastic-net regularization to learn the output layer weights for the ELM. Accordingly, efficient algorithm is developed to optimize the sparse representation problem. The elastic-net regularization is expected to learn more selective and compact solutions for the ELM, thus it is beneficial for feature learning. Compared with traditional deep learning methods, the proposed algorithm can achieve better classification accuracy while cost less computational time.Lastly, to solve the problems caused by lacking of labeled training data, we propose a discriminative semi-supervised sparse representation algorithm with graph regularization for image classification. The algorithm can learn labels for the unlabeled data through utilizing the information included in the labeled and unlabeled data. The learned labels not only keep the same manifold structure as the original data, but also pose excellent discriminative ability. To fully exploit the underlying information contained in the unla-beled data, we define a new kind of within-class affinity matrix and between-class affinity matrix by designating extra weights to the data within the neighborhood. As a result, discriminative features can be extracted from the data. To deal with the data that cannot be linearly classified, we also propose a kernel sparse representation algorithm which can linearly classify the data in high dimensional space. In addition, we develop two efficient algorithms to optimize the two sparse representation problems respectively. Experimental results demonstrate that the proposed algorithms can achieve excellent performance when the labeled training data are scarce.