Fast online learning algorithm for landmark recognition based on BoW framework

Abstract

In this paper, we propose a fast online learning framework for landmark recognition based on single hidden layer feedforward neural networks (SLFNs). Conventional landmark recognition frameworks generally assume that all images are available at hand to train the classifier. However, in real world applications, people may encounter the issue that the classifier built on the existing landmark dataset needs to be tuned when new landmark images are collected. To address this issue, a fast online sequential learning framework based on the recent extreme learning machine (ELM) which can update the classifier by learning the new images one-by-one or chunk-by-chunk is developed for the landmark recognition. The recent spatial pyramid kernel bag-of-words (BoW) method is employed for the feature extraction of landmark images. To show the effectiveness of the proposed online learning framework, the batch mode learning method based on ELM is also employed for comparison. Experimental results based on the landmark database collected from the campus in Nanyang Technological University (NTU) are also given to verify our proposed online learning framework.