Dimension reduction based SPM for image classification

It is difficult to classify images with high accuracy when the dataset is relatively large. We try to improve classification precision in spatial pyramid matching (SPM) framework by dimension reduction. When clustering high dimension data, researchers encounter dimensional curse problem which would weaken statistical significance of the data. This problem degrades the performance of SPM and other related works based on clustering property of the high dimensional ”SIFT” features. We propose a global dimensional reduction approach, reducing 128-d SIFT features to 32d, and follow processes of locality-constrained linear coding to calculate feature histogram. Experimental results show that the proposed method leads to a better clustering property for the local descriptors, and increases image classification precision comparing to other state of the art algorithms on several image datasets.Bosch2007Bosch2007.