Adaptive Techniques for Content Based Image Retrieval (CBIR): Study and Applications
Loading...
Date
2021-04-04
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Skikda - 20 Août 1955
Abstract
The significant growth of images repositories in terms of content and usage has increased the
need of conceiving powerful image retrieval systems. In this regard, CBIR is the process of
searching images given a query based on their visual aspects.
One of the most serious challenges that is facing CBIR systems is “the semantic gap”. This
gap occurs between images representation given by a CBIR system and the true semantics
included within these images. The semantic gap should be bridged or at least reduced in order
to bring closer the low level vision of the CBIR system and the high level vision of the Human
Visual System (HVS). This could be achieved through designing CBIR systems that behave
with respect to the human understanding of images.
In this thesis, we address this challenge by shedding light on a paradigm that has already imposed
itself in many fields, that is the adaptation paradigm. Indeed, there is a need to inject a flexibility
in the designed CBIR systems, so that they behave depending on some criterion, mainly the
input data. We particularly integrate the adaptation into images characterization phase so as
to reduce the semantic gap. This adaptation is carried out by selecting the characterization
algorithm or tuning its parameters in a training phase, in order to increase its performance with
respect to the target task.
First, we start with the selection paradigm, where we apply the algorithm selection based
on RICE model to select the best performing CBIR-algorithm for each image. In this work,
the CBIR-algorithm was decomposed to many variants, where the feature variant was selected
based on the input image. In the second contribution, we propose to use a Convolutional
Neural Network (CNN) in the feature extraction stage. This latter is first finetuned on the
target images dataset in order to adapt its parameters. The resulting network is then used to
extract the class probability vectors from input images to use them as representations. Last, our
main contribution consists of using an ensemble of CNNs that collaborate to extract relevant
features from images. Herein, the class probability vectors are extracted from images using each
ensemble member. They are then combined to make a powerful image representation.
Finally, it is noteworthy that all our contributions have led to a significant improvement in the
performance of the developed CBIR systems. This latter is a relevant indicator of narrowing
the semantic gap within these systems, which is the ultimate goal of this thesis