Feature coding for image classification based on saliency detection and fuzzy reasoning and its application in elevator videos

Feature coding is an fundamental step in bag-of-words based model for image classification and have drawn increasing attention in recent works. However, there still exits ambiguity problem, and it is also sensitiveness to unusual features. To improve the stability and robustness, we introduce saliency detection and fuzzy reasoning rules to propose an novel coding scheme. In detail, saliency maps generated by saliency detection are first used to divide each image into salient and non-salient region, then a structured dictionary is obtained by combing two separated codebooks in them. Secondly, fuzzy reasoning rules are introduced to choose the most salient and stable codewords to encode. Finally, saliency maps are incorporated into pooling operation named saliency based spatial pooling to introduce spatial information. Experiments on several datasets demonstrate our approach outperforms all other coding methods in image classification. Furthermore, we also apply it into elevator video event classification, which shows the potential application in intelligent elevator video surveillance, such as overload detection, violence detection, video summarization. Key-Words: Image classification, feature coding, saliency detection, fuzzy reasoning, elevator video event * Corresponding author.