Due to the coarse fineness of global Gist features in scene categorization tasks

we propose a local Gist feature description based on a dense grid. It uses a spatial pyramid structure to add distribution information and introduces the RGB color space to add color information. The feature matching process is kernelized by an efficient match kernel which mea-sures the similarity between local features based on the BOW model. The scene categorization task can be done with linear SVM. Experiment shows the influence to the classification accuracy with local Gist features which have different scale

orientation

fineness

match kernels and numbers of training samples. By using the classification result of the global Gist feature and dense SIFT features on the OT scene dataset

we demonstrate that the proposed feature construction method and classification model are efficient.