ELSE: Ensemble Learning System with Evolution for Content Based Image Retrieva

Abstract

Images and graphics are among the most important media formats for human communication and they provide a rich amount of information for people to understand the world. With the rapid development of digital imaging techniques and Internet, more and more images are available to public. Consequently, there is an increasingly high demand for effective and efficient image indexing and retrieval methods. However with the widely spread digital imaging devices, textual annotation of images be- comes impractical and inefficient for image representation and retrieval. To diminish the reliance on the textual annotations and associated meta- data for image search, the content based image retrieval (CBIR) has be- come one of the most popular topics in the field of computer vision and pattern recognition. In CBIR, the image representations are generated through the visual clues like color, texture, or shape of objects; and cer- tain machine learning algorithms are applied to understand the image semantics for meaningful image retrieval. However, despite the great deal of research work, the image retrieval performance of the CBIR sys- tems is not satisfactory due to the existent semantic gap between the low-level image representations and high-level visual concepts. To bridge this gap to some extent, three major issues in the active field of CBIR are investigated in this thesis, that are: consistency enhancement during the semantic association, improvement in the relevance feedback (RF) mechanism, and generation of a stable semantic classifier. Consistency enhancement in semantic association process, addresses the two main reasons, due to which the conventional CBIR systems are not able to produce the effective retrieval results. These are: the lack of output verification and neighborhood similarity avoidance. Due to these problems the image response is very inconsistent and the target output contains far more wrong results as compared to the right results. In this thesis, we concentrate these issues by applying the Neural Networks over the bag of images, and exploring the query’s semantic association space. In this regard semantic response of the top query neighbors is also taken into the account. The potential image retrieval is strongly dependent on the efficacy of the image representations. Therefore the deep texture analysis is performed through the best basis of the wavelet packets and Gabor filter to explore the representations which may serve as the most effective basis for automatic image retrieval. The Relevance feedback (RF) in CBIR, specifically focuses on the cus- tomization of the search results to the user’s query preferences based on the several feedback rounds. These systems can easily be mislead by theover-sensitivity in the subjective labeling. Another problem that usu- ally occur is the imbalanced class distribution that makes the classifier learning a real challenge. The amalgamation of both is a big reason for the user frustration, and hence make the system of no practical use. We overcome both of these issues through Genetic Algorithms, and demon- strated the positive performance impacts by SVM classifier. Extending the ideas for imbalance distribution in binary classification to multi-category environment leads in the form of a stable semantic classi- fier. The semantic association becomes even more challenging when there are many categories enrolled. The reason is that: the positive training samples for a particular class are naturally far less then the training samples from many other classes. Weak classifiers like SVM and Neural networks are not able to perform well in these circumstances. Therefore the most effective solution lies in the exploitation of the combined basis function for these week candidates. The Genetic classifier comity learn- ing (GCCL) is tuned for overcoming the limitations like classification biasness in multi-category environment, incompatible parameter estima- tion, and overfitting due to the high dimensional nature of the feature vectors compare to the training sets. The qualitative and quantitative analysis shows that the proposed method outperform many state-of-the- art methods.