Semi-supervised Image Ification Based On Sparse Coding Spatial Pyramid Matching

With the development of multi-media technology, a variety of images are emerging in multitude nowadays. As the basic method of image data organization, image classification is always the hotspot of research area. The task of image classification mainly contained two steps:(a)Firstly, we have to represent the visual image into numeric information, i.e. image representation.(b)Then giving the specific image representation, we have to choose the reasonable image classification model to train until we obtain the image classification model. Both of the two steps have a great important to the performance of image classification, and they do supplement each other.Staring from this point, this paper spreads the research of image classification. First we apply the sparse coding and spatial pyramid matching model to obtain the image representation. Through extracting the image’s SIFT feature to train a overcomplete dictionary, we use the learned dictionary to do sparse coding for the SIFT features. After obtaining the codes of SIFT features, we then combine the spatial pyramid matching model to get the final representation of the image. Based on the image representation, we can construct a linear Mercer kernel (like the linear Support Vector Machine)to do the image classification task, which leads to the O(n) computation complexity in the training phase, and constant complexity in the testing phase. The traditional method of image representation will result in a nonlinear Mercer kernel to ensure a good performance. Classification model with nonlinear kernel, like nonlinear SVM, will have O(n3) computation complexity and O(n2) space complexity in the training phase, and O(n) computation complexity in the prediction phase. This scalability of the traditional image representation method implies a severe limitation—it is nontrivial to apply them to real word applications,whose training size is typically far beyond thousands.Because the image representation obtained by sparse coding spatial pyramid matching model is sparse and high dimension, it is reasonable to assume that these vectors lie on a low dimension sub-manifold. Based on this assumption, we adopt the manifold regularization semi-supervised learning model as the classifier, which incorporates the information of unlabeled data into the decision function using the manifold regularization. By doing this, the classifier try to exploit the intrinsic data structure to obtain good performance. In practical systems, more unlabeled data is available than labeled data. Semi-supervised method is able to make better use of unlabeled data to improve recognition performance. In combination with the two steps, this paper proposes a LapSVM semi-supervised image classification based on sparse coding spatial pyramid matching. The experiment shows the efficiency of our method.