In this paper

we present a new scene categorization algorithm based on supervised subspace modeling and sparse representation. The proposed method implements supervised dictionary learning via decomposing the unsupervised sparse dictionary learning model into a group of independent optimization problems. After learning the dictionaries of all categories

we aggregate them to form a global dictionary and encode each local feature of an image based on it. After using spatial pyramid representation and max pooling of local features’ coding vectors

the final holistic feature depicting a scene image can be retrived. Comprehensive experimental results on four popular benchmark scene datasets show that our method achieves very promising result compared to existing state-of-the-art techniques.