Shallow Feature Based Image-Text Retrieval Using Kernel Extreme Learning Machine

Image-text retrieval has become more popular day-by-day due to its huge application on Facebook, YouTube, Google, LinkedIn, Twitter, etc. It is a cross-modal retrieval system considered a multi-class problem, which presents a challenging task in computer vision. Finding an appropriate combination of feature extraction and classification models is necessary to solve this multi-class problem effectively. This paper proposes a shallow feature-based cross-modal information retrieval system. Specifically, the scale-invariant feature transform (SIFT) is utilized as a shallow feature descriptor and the kernel extreme learning machine (KELM) is adopted as the classification model. Semantic matching (SM), correlation matching (CM), and semantic correlation matching (SCM) are applied in the matching strategy. The publicly available Wiki database is used for model learning. The mean average precision (mAP) is calculated as a statistical measure for each matching strategy, utilizing various distance metrics. The experimental results demonstrate that KELM achieves the highest mAP for SM, using a centered correlation measure, outperforming other matching strategies and statistical measures. Additionally, the results reveal that the SIFT feature-based KELM combination performs better than other competitive shallow feature-based approaches.