Research On Image Classification Algorithm Based On Subspace Learning And Sparse Coding

With the rapid development of multimedia and internet technology, image information presents explosive growth. It has become an urgent problem to quickly seek the demand data from the massive data. In order to meet people’s requirements, the image classification technology can automatically retrieve the image by analyzing and processing the image data. Thus, it has higher theoretical and practical value on the study of image classification.In the dissertation, we use SIFT feature and the pixel feature of the image, and carry on a detailed and depth study on the subject of image classification from the perspectives of high-dimensional data reduction, optimization subspace and reconstruction error classification respectively. Finally we also make some research results. The main works of this dissertation are as follows:(1) For the image processing, the block image will be better to use spatial information of each pixel to obtain more details, and the block image may be more suitable for parallel computing on the computer. Thus, we propose to process the image data into blocks to resolve the problem and classify in subspace. However, the traditional mathematic theory cannot derivate the subspace model for the block subspace, we propose a subspace optimization model based on the particle swarm algorithm. We use the particle swarm algorithm to optimize the block subspace, and we ultimately get an optimized subspace to do image classification.(2) The classification algorithm that is used to process these high-dimensional data often suffers from the low accuracy and high computational complexity. In order to overcome these problems, we propose a framework of transforming images from a high-dimensional image space to a low-dimensional target image space, based on learning an orthogonal smooth subspace. Firstly, we extract SIFT feature, and a sparse coding followed by spatial pyramid max pooling is used to get a high-dimensional descriptor for image. Then, the high-dimensional image descriptor is mapped into an orthonormal and smooth subspace to reduce dimension. The experiments on the image classification verify the effectiveness of the algorithm.(3) The image classification algorithm based on sparse coding cannot take full advantage of the category information of the selected training data, which also makes the dictionary is unsupervised. In order to make the best of known prior information, we propose a supervised reconstruction error classification algorithm. We will add the lable information to dictionary to obtain effective dictionary description in the process, and we use the minimization criterion of reconstruction error to do image classification. The experiments on the image classification also verify that the algorithm can effectively improve the classification performance.