SSVR-based image semantic retrieval

To bridge the wide semantic gap between the image low-level visual features and the high-level concept conveyed from images is still a challenging job. In this paper, an image semantic representation model (ISRM) was proposed based on statistical learning theory and smooth support vector regression (SSVR). This model is a five-tuple which consists of primitive image set, image feature set, image semantic set, semantic rule set and semantic mappings. The example-based high-level semantic retrieval algorithm (EHSR) and the text-based high-level semantic retrieval algorithm (THSR) for image retrieval using high-level semantic content were designed and implemented respectively. The performance of an experimental image retrieval system constructed according to aforementioned approaches was evaluated on a database of around 3000 images. The experimental results show that ISRM model and EHSR and THSR algorithms are effective in describing image high-level semantic content and can provide flexible and efficient image retrieval performance.