MS Thesis Defense: Mohbat Tharani

Challenging the dogma of Relevance Feedback in Content based Image Retrieval Systems with Deep Learning

May, 2018 at 3:00 pm in Smart Room 9-105 SBASSE.

Abstract

Association of images to their content based similar images in a database, is quite a fascinating challenge specially on social media platform where billions of tagged and untagged images are uploaded by random users every month. This task could be solved using deep neural perception networks which successfully acquired remarkable popularity in a number of diverse fields especially image classification and scene understanding. Inspired by the idea, we have developed a residual encoder-decoder visual descriptors and a novel relevance feedback-less CBIR framework, employing residual discriminative network replacing conventional similarity matrix computation in the CBIR systems for remote-sensing images. We have evaluated the performance of our approach on two publicly available benchmark datasets, namely University of California (UC) Merced Land Use/Land Cover dataset and High-resolution Satellite Scene dataset (SceneSet). With experimentation we prove that our methodology performs better than all the previous retrieval methods used for remote-sensing images employing hand crafted and deep/CNN visual descriptors with traditional numerical distance and relevance feedback.