Skip to Main Content
We introduce a new approach into scene classification problem related to Bag-of-Words (BoW) representation. Category specific filter banks are generated on distinctive feature channels with varying scales by using Graph-Based Visual Saliency (GBVS) algorithm. After preprocessing each image using filter banks, dense Scale Invariant Feature Transform (SIFT) method is applied to the filtered samples at regular spacing grids. In order to achieve scale invariancy, we concatenate SIFT-like descriptors from filtered images of different scales within visual channels. In image representation stage, BoW modeling is improved by adding spatial information and a probabilistic voting scheme. We compare the proposed algorithm with the most promising methods in the literature, using a very challenging and popular 15-class-dataset. It is seen in experiments that our method noticeably outperforms the others.