Page 152 - 2023-Vol19-Issue2

148 |
ters of the model are tuned using the Stochastic Gradient De-
scent (SGD) optimizer to optimize retrieval performance. The
similarity between images is measured using the Manhattan
distance metric, enabling the retrieval of highly similar images high dimensionality and reduced retrieval accuracy [17–21].
from the database. The DTLDN-CBIRA technique demon-
strates its novelty in the design of the plant disease image
retrieval process. The performance of the DTLDN-CBIRA
model is evaluated using a benchmark dataset. The results
highlight the superiority of the DTLDN-CBIRA model over
recent methods, achieving a maximum precision of 100Au-
thors in [28] proposed an innovative approach to CBIR, a
technique vital for finding images within expansive, unlabeled
image collections. The authors recognized the importance of
similarity computations and feature representation in ensur-
ing the effectiveness of a CBIR system. Key image features
such as color, shape, texture, and gradient were acknowledged
as essential elements in image representation. A Local Bi-
nary Pattern (LBP), an efficient yet straightforward texture
descriptor, was applied to label image pixels by thresholding
the neighborhood of each pixel and interpreting the result
as a binary number. Additionally, they presented a noise-
robust binary pattern known as the ’Median Binary Pattern’.
When applied to a practical dataset named CBIR 50, their
method yielded encouraging results. Compared to existing
approaches, the proposed method attained an Average Recov-
ery Precision (ARP) and an Average Recovery Rate (ARR)
of 68.1% and 33.55%, respectively, employing Noise Robust
Binary Patterns. This work constitutes a crucial component of
the ongoing discourse on enhancing CBIR efficiency, demon-
strating that comprehensive feature representation can sig-
nificantly improve image retrieval outcomes. The authors
introduced in this paper [29] an entropy-based measure that
considers the grouping property of returned relevant images,
which is essential for fast exploration of results through user
visual inspection. They emphasized that common evalua-
tion measures do not illustrate the grouping property of the
returned relevant images and miss the interrelation between
them. The proposed performance measure is described as easy
to understand and implement, and its discriminating power
is demonstrated through a comparative study with existing
CBIR evaluation measures. This paper contributes to the field
by addressing the limitations of standard measures, especially
for image retrieval, and by extending the evaluation scale to
achieve better discriminating power. This allows for different
evaluations of two systems that have the same precision value.
In this work, we aimed to tackle some of the inherent limita-
tions in the field of CBIR with a focus on the use of DL. Prior
research has shown that while DL models, such as CNNs,
have been successful in scene categorization, they are not
without their flaws. These issues include a complicated train-
ing procedure, a need for vast amounts of training data, and
Mohammed, Oraibi & Hussain
extended training times [15, 22]. Some existing CBIR models,
such as the BoVF and OB, have also overlooked critical infor-
mation like spatial and semantic data, leading to issues like
Our work builds upon these existing methods and proposes a
novel CBIR approach that not only addresses the limitations
of previous work but also enhances image retrieval accuracy.
We employ a hard voting ensemble approach to aggregate fea-
tures extracted from three potent DL architectures: Inception,
Exception, and Mobilenet. This ensemble strategy allows us
to bridge the semantic gap between low-level image features
and human visual perception, resulting in a more accurate
and effective image retrieval process. One key strength of our
method is that it bypasses the complex training process and
extensive data requirements typical of CNN-based DL models.
By using an ensemble of pre-trained models, we effectively
utilize their collective strengths, enhancing the robustness and
accuracy of our CBIR system without the need for exhaustive
training. Our method also addresses the neglect of spatial
and semantic information in existing models. The Inception,
Exception, and Mobilenet architectures each incorporate tech-
niques for capturing these types of data, contributing to a
more comprehensive feature extraction process. By harness-
ing these architectures in tandem, we effectively capture a
broader and richer set of image features. In conclusion, by
building on the strengths of robust DL architectures and ad-
dressing the shortcomings of traditional CBIR approaches,
our ensemble-based method presents a powerful, effective,
and efficient solution for image retrieval. It moves a step
closer to bridging the gap between low-level image features
and human visual perception, making substantial strides in
the development of advanced CBIR systems.
III. CBIR SYSTEM
A. Overview of the General Flowchart
The platform of the CBIR system [1], which may be further
separated into an offline and an online subsystem, is intro-
duced in this section. This is represented in Figure 1. Each
block in the off-line subsystem has an index in the retrieval
database that is coded by the extracted feature vector from
the image. Following the input of a query picture, the feature
vector of that image is extracted in the online subsystem using
the same method as the feature vectors of the photos in the re-
trieval dataset. Once all potential photos in the database have
been scored using a similarity metric, this feature vector will
be used. Images that score over a certain threshold are chosen
to be further refined by increasing the visual context relative to
the initial query. The retrieval system’s outputs or results that
are probability-ordered are these pictures that are arranged in
ascending order of the re-rank score. For the dataset indexing
in this system, which uses a specific similarity metric, the