A novel weight assignment based Image retrieval using bovw model And deep hashing techniques

Abstract

Visual information is available in abundance and it has been constantly increasing due to the present Internet and digital advancements. ccessing images like filtering, browsing, retrieving, and classifying are Retrieval is a fast-growing research field that incorporates cross-disciplinary features like Information Retrieval, Machine Learning, and Computer Vision. At the earlier stages of image retrieval, Text Based Image Retrieval (TBIR) requires meta-data in textual format to retrieve images for textual queries. It functions well as long as the images are meaningfully tagged. But, its limitations are an increase in the manual annotation that involves human experts and accuracy obtained that are subjected to the human annotations. These issues create a necessity for Content Based Image Retrieval (CBIR). It uses visual content to describe images. Three research works have been proposed in this thesis concerning upgrading the CBIR systems. The first work is an extension work of the Bag of Visual Words (BoVW) model, where BoVW is a widely recognized method to address the semantic gap problem existing in CBIR. Despite its ample acceptance, it suffers from low discrimination ability among visual features and lacks spatial information due to order-less visual words. To improve the discrimination ability from the generated visual words, important visual words are to be identified based on their contents. These important visual words for a class are designed as Visual Patterns. Visual Patterns are the collection of important and unique visual words contributing to each class. They are determined by the weights of the visual words calculated based on their information richness from all the images belonging to each class newline