Content-based image retrieval based on indexing of code words and metadata attributes in large database

The use of local feature extractor algorithms for content-based image retrieval systems for large data sets in real-world applications is not appropriate in terms of time, cost and functionality. The growth of the feature set yields to prolong duration of process since all features in the set have to be compared to each other. In this study code book technique has been utilized and enhanced in a content-based image retrieval system for the real-world applications. Obtained local features of images were quantized into a new space that represents all the features by the help of cluster. Those cluster centers were used like in word indices in books. Hadoop framework is used for high amounts of resource requirements in generating code book and code words, thus a feasible system has been designed. With the help of this technique, multi-dimensional vector space of feature set were reduced into one dimension and translated into visual words, the latter was indexed using reverse indexing algorithm therefore image retrieval time is shortened. Since search was conducted faster in one dimension, it has been shown that, running multiple algorithms collectively was enabled. In this study, variety of algorithms and features have been analyzed, local feature extractor algorithms such as MSER, SIFT and SURF algorithms were used. In addition to these known local feature extraction methods, using general feature extraction algorithms such as PHOG, LBP, Tamura, RILBP methods applied to local key points of SURF algorithm in SURF window size in order to obtain localized general feature of images. These algorithms applied to SURF window width subsequently applying these technique to different SURF window width showed that different feature sets could be obtained. Therefore, the method was applied to the optimum window size yielded to the best feature sets. All general feature extraction algorithms applied to new window size to derive new features, than those new features were indexed for later retrieval. With the gain in retrieval speed and diversity of features, accuracy increased using multiple algorithms over one algorithm. A system has been designed where visual words and image labels can be indexed and queried.