Boosting classifiers for scene category recognition
Fuxiang Lu, Jun Huang, Kun Zhan · Lanzhou University Institutional Repository · 2015
This paper presents a method for recognizing natural scene categories based not only on approximate global geometric correspondence between the features of the same kind, but also on complementary information cues offered by heterogeneous features. This technique works by dividing the image into increasingly fine sub-cells and computing the bag-of-features found inside each sub-cell. The discriminative power of each resulting spatial pyramid depends largely on its specific choices on interest point detector and local region descriptor involved in computing the bag-of-features. Different choices on interest point detector and local region descriptor lead to a powerful image representation: multiple pyramid histograms of words (mPHOW), which is a simple and computationally efficient extension of pyramid histogram of words (PHOW). In order to recognize an unknown image as correctly as possible, this paper first employs multi-class support vector machine (SVM) classifiers to compute posterior probabilities from the individual PHOWs, and then adopt the boosting algorithm to combine the variants of SVM, each trained on a single PHOW, to obtain the improved estimate of the “final” posterior probabilities. Our proposed method is evaluated on three benchmark scene datasets: OT, FP, and LSP. Results demonstrate that the proposed method outperforms the compared algorithms consistently. © 2015 ISSN.