Research On Image Representation And Classification Based On Non-negative Matrix Factorization

Recently with the rapid development of multimedia technologies it becomes much easierto obtain high quality images, how to represent and classify these high quality images hasbecome a hot research topic. On one hand, high quality images with high dimensionalitypossess more plentiful features, while on the other hand high-dimensionality makes it difficultfor us to address data, such high-dimensionality data good for image representation usuallylead to “curse of dimensionality”. Therefore, it is necessary to conduct dimensionalityreduction before image classification.At present, there exist a variety of methods for data dimension reduction. However,among the obtained results through such methods, some values are negative, in which case itcan not been endowed physical significance in practical applications obviously, which is notexpected to appear in real-world tasks. While nonnegative matrix factorization (NMF) methodcan overtake the above issue and guarantee that the data values after factorization are allnon-negative. The data processed by NMF is purely additive and sparse to some degree. Pureadditive shows the reasonability of factorization while sparseness can inhibit the outside fromaffecting data features, and make the data robust. At the same time simple but effectiveiterative updating algorithm is adopted when factorizing non-negative data, information aboutpart-based features can be finally obtained by continuous learning, which caters to thecognitive process of people’s recognizing things from part to whole.Image representation and classification are very important issues in pattern recognitionarea. During the process of image classification, whether the feature extraction is reasonableand whether the classification function is optimal will directly affect the image classificationresults. In this paper we apply nonnegative matrix factorization algorithm to image datadimension reduction and local feature extraction and classification. In this paper, we studysparseness-based constrained NMF, graph regularized and label-based constrained NMF interms of image processing field. Synthesizing such theories, we put forward three improvesNMF algorithms, including: constrained NMF with sparseness, graph regularized and sparseNMF with hard constraints, supervised constrained NMF with sparseness. Afterdimensionality reduction and feature extraction, we utilize K-means for clustering. In order toguarantee that the supervised NMF algorithm has better abilities of classification, weincorporate support vector machine theory into classification. Extensive experiments arecarried out on common facial and objective databases, and experimental results demonstratethe reasonability and effectiveness of our proposed algorithm here.