Asian Research Thesis Index

Abstract

Increased availability of devices with camera and rapid growth and development of the internet has enlarged image databases for individuals and organizations. It has raised a need for an effective and efficient approach for searching, browsing and retrieving images from large image databases of different fields of life e.g. medicine, economics, education, architecture, etc. This problem is being addressed using Content-Based Image Retrieval (CBIR) by developing the number of algorithms that use different kind of features that includes both low and high level features. These features are used to measure the similarity between images. But these features extraction and similarity measure approaches have many limitations that resulted in excessive response time and retrieval accuracy. CBIR mainly depends on extracted features from image contents and if possible, with semantic concepts. So, the semantic gap needs to be narrowed down while extracting features that can help in increasing retrieval accuracy. To minimize the retrieval time, instead of comparing features, we can transform features into deep hash codes. Comparing numbers takes less time than text or features. So, we can reduce the retrieval time by comparing the deep hash codes of query and database images. We proposed changes in the objective function of Convolutional Auto encoder to extract features while triplet loss function in the last layer to convert features to binary hash code with semantic preservation and similar code for same types of extracted features. This approach reduces the dimensions of the features to binary hash code length that make retrieval easier and faster. Proposed image retrieval system has been tested with different data sets and comparisons have been made with different loss functions and hashing approaches. Experimental results indicate the effectiveness of the proposed approach and show improved retrieval accuracy.

Add/Update Thesis