Design of Efficient Methods for Content Based Image and Video Classification and Retrieval

Abstract

Images and videos are the most preferred source of information used by people around the world. As they can be easily captured and uploaded on internet, these repositories are increasing every second. This brings the major challenge to access, analyse and retrieve them. Powerful search techniques are needed to access these data. Text based search is inefficient and impractical. Retrieval of images and videos through the analysis of their visual content is therefore an exciting and a worthwhile research challenge. Content based search is a powerful and automated technique to find the required image and video in large databases. However, its accuracy is highly dependent on features used for retrieval. Low level visual features like color, texture, edge, shape etc. do not directly capture the high level semantics or contextual information present in the images. This mismatch or gap is termed as semantic gap . Proposing effective feature descriptor to bridge this gap is an active research topic in image and video processing. The feature descriptor can be designed manually or in an automatic manner such that there is balance between accuracy and computational complexity. newlineThis thesis is dedicated to the design of efficient methods for content based image and video classification and retrieval (CBIR and CBVR). Effective CBIR systems using i) Fusion of hand crafted features ii) Deep features and iii) Combination of hand crafted and deep features are proposed in this work. We have also proposed a CBVR system which involves key frame selection and feature extraction using hand crafted and deep learning techniques. newlineAn effective feature descriptor for CBIR by fusion of color, texture and edge features is proposed. Color Coherence Vector (CCV), Center Symmetric Local Binary Pattern (CSLBP) and Edge Histogram Descriptor (EHD) are used for extracting color, texture and edge feature of images respectively. Concise representation of color distribution in an image, capturing both frequency and spatial coherence of colors is obtai