Article Preview
TopIntroduction
Current state-of-the-art image retrieval applications rely on metadata to recognize the content of images. Source of metadata information usually is manual annotation, a process that is prone to inaccuracy and cannot scale with the continually growing size of today’s multimedia content. Automated annotation requires scene analysis and object recognition using information obtained from acquisition system.
Scene understanding has been studied for decades in specialized areas such as surveillance and sporting event recognition. Main objective of this field is to provide more precise and accurate description of the scene to better serve the user queries. A successful solution for this task must be able to recognize the scene elements and to be able to interpret various interactions among them such as events.
Image visual features and the contextual information are two main sources of information used by visual cognitive system in humans which performs scene recognition. Variations in image quality or acquisition conditions can significantly affect the quality of the visual features. To mitigate variations and improve robustness of features, algorithms such as Scale-Invariant Feature Transform (SIFT) (Lowe, 2004), Speed-Up Robust Features (SURF) (Bay, Ess, Tuytelaars, & Van Gool, 2008), Binary Robust Invariant Scalable Keypoints (BRISK) (Stefan Leutenegger & Siegwart, 2011) and Oriented Fast and Rotated Brief (ORB) (Rublee, Rabaud, Konolige, & Bradski, 2011) have been developed. Starting with SIFT as first generation of gradient based methods, these algorithms detect best keypoints in a scale pyramid type and extract orientation using the directed gradients or moments. The computational costs of the feature point calculation in earlier methods have been the motivation for continuous development of more efficient and robust algorithms. But even most superior algorithms fail to produce high quality features in extremely poor and degraded image conditions. Study on human visual cognitive system show superior recognition ability even in children when images of various qualities are presented with a familiar “context”. In computer vision, the context is referred to any information that can be used in accurate semantical understanding of a scene and recognition of its element.
Previous works show evidence that performance of recognition systems can be improved when context is exploited (Wang, Chen, & Wu, 2011; Zhang, Kalashnikov, Mehrotra, & Vaisenberg, 2014; Galleguillos & Belongie, 2010). Contextual models can capture properties, relationships and interactions among image components to infer more accurate meaning of a scene (Choi, Torralba, & Willsky, 2011; Zhu, Nayak, & Roy-Chowdhury, 2013).
One of the most effective context taxonomies is semantic context which is expressed as co-occurrence statistics of image components (Shotton, Winn, Rother, & Criminisi, 2009). In natural language analysis, “Latent Dirichlet Allocation” (LDA) is a generative process that uses latent variables in discovery of intrinsic structure of document and maximizes words co-appearance probability. Given a set of documents as collections of words, LDA decomposes distribution of the words in each document into the mixture of K topics such that a document is viewed as the mixture of topics which itself is multinomial distribution of words. Supervised LDA learns a small set of topics that account for most of the correlation and reveals semantic structure of the document. These models explore various associations between the image features and the corresponding annotation words (Wang, Blei, & Fei-Fei, 2009). LDA topic modeling approach has been widely used in computer vision to discover latent semantical correlation of image visual patterns for clustering and classification applications (Huang, Zhou, & Zhang, 2014; Bahmanyar, Cui, & Datcu, 2015).