Rotation, Translation, and Scale Invariant Bag of Feature Based on Feature Density

Abstract

In this paper, we propose a feature representation that achieves translation, rotation, and scale invariant simultaneously. We first proposed a novel component, called Block Based Integral Image, to search the densest region of feature points. This aims to find the center of potential object in the image. Then, with the improved object center, we apply Spatial Pyramid Ring (SPR) by to handle translation and rotation invariant representation. After that, histogram equalization technique is utilized to adjust representation for scale invariant. The experimental results are demonstrated on different datasets by image classification task. Experimental results show that our translation, rotation, and scale invariant representation achieves higher accuracy than the previous methods.