Scene classification based on SIFT combined with GIST

To overcome the disadvantage of global GIST features that lose local information needed for scene classification tasks, a new scene feature description method that combines global GIST with local SIFT features is proposed in this paper. Firstly, local context information and global RGB color quantization information are introduced into the traditional SIFT and GIST features respectively, and then the similarity between the characteristics of the scene is measured based on BOW (Bag Of Words). Finally, the scene classification task is performed with SVM. The influence on classification accuracy of the combined features with different SVM match kernels and BOW is investigated in experiment, and based on three scene datasets, the classification results of the combined feature are compared with that of the methods in literature based on single feature of global GIST or local SIFT, the experimental results show the efficiency of the proposed feature construction method.