Image Classification Based On Sparse Coding

Image classification is a fundamental research among computer vision, machine learning, pattern recognition, which plays an important role in the application domain including visional information retrieval, data filtering, medical image recognition and so on. Based on the different definition of classification function, the traditional image classification methods include Bayesian classification, Fisher classification and SVM classification. According to that whether or not the training data is supplied with labels, the general image classification methods are divided into supervised classification, semi-supervised classification and unsupervised classification.As is known, sparse coding is one of the artificial neutral networks, which simulates mammalian visual cortex’s reaction to the signal stimulator. It is extensive applied in the signal processing. Recently, sparse coding has been extended in the computer vision, dealing with the bottom visual processing tasks such as image denoising, image to defect and image restoration. For the reason that sparse coding is a method that follows the humans visual reaction law, it is introduced to solve high-level visual processing tasks, including image classification. Compared to the traditional image classification methods, spare coding is a totally new methods, imperfect and a lot of work should be put into its study.In this paper, first we review the traditional image classification methods and analyze the advantages and disadvantages of each of them and suitable application scene, respectively. Afterwards, we summarize the history of sparse coding and the present research state, and make a deep study among some of the sparse coding methods. We summarize the sparse coding in mathematics. Combining with the new image descriptor and image representation methods, we could improve the image classification accuracy.In this paper, we have made a combination between NCUT cluster and sparse coding, which makes the classification tasks evolves from coarse level to precise level. At the same time, we have introduced the up-to-date image descriptor and image representation method, which improves the classification accuracy more. We can get the conclusion that our method makes sense and does well according to the results of our experiments.