Page 150 - 2023-Vol19-Issue2

146 |
two components. In real life, retrieving an exact picture from
a sizable database is still difficult. The biggest problem is
the semantic mismatch between the image’s low-level visual
qualities and its high-level meaning [2]. This gap has been the
subject of countless research during the last three decades [3]. extensively employed in existing CBIR techniques, it neglects
There are several ways to translate high-level concepts in pic-
tures into features. The basis of CBIR is comprised of these
elements. According to the methodologies used for feature
extraction, global and local characteristics are two common
categories for features. Global characteristics of the image,
including color, texture, shape, and spatial details, serve as a to a large dimensionality difficulty when applied. This high di-
depiction of the entire item. They benefit from being quicker
at feature extraction and similarity calculations [4]. On the
other hand, they fail to recognize the difference between the
image’s backdrop and the item in it (different image parts).
They are therefore inappropriate for object identification or
retrieval in complicated settings [5]. However, they are accept- amounts of training data, and excessive training time are sig-
able for object categorization and detection [6]. There have
been significant attempts made by academia and industry to
close this semantic gap. As a result, CBIR has been shown to
make significant progress recently. For instance, well-known
search engines like Google and Baidu can look for similar im-
ages for any image. Several e-commerce websites, including
Alibaba, Amazon, and eBay provide comparable commodi-
ties search features. The content suggestion features on social
media networks like Pinterest are comparable [1].
Query By Image Content (QBIC) and CBIR are related
by nature [7]. Early in the 1990s, CBIR was founded [8].
This automated process uses a picture as a query to present
a collection of photos that correspond to the query. The low-
level picture attributes, such as texture, color, and shape, are into the following sections: Section II provides an overview
taken from the database images in order to categorize them.
We assume that images in the same category will share similar
traits. Retrieval of images will therefore see an incredible
increase in efficiency when similarity measurement is carried
out based on picture attributes [9]. One of the subcategories
of the soft computing phenomena known as Deep Learning
(DL) which allows for the retrieval of data from millions
of separated pictures [10]. A content-based picture retrieval
system performs optimally when the feature representation
and similarity evaluation, which have been extensively studied
by multimedia researchers for decades, are used. Even though
several solutions have been proposed, it is still among the
trickiest issues in CBIR research. This challenge can be linked
to the core challenge in AI: how to build and train AI tools
that can carry out routine human tasks [11, 12].
The field of CBIR faces several significant challenges that
impede the development of efficient and accurate retrieval
systems. One of the primary issues is the semantic gap be-
tween low-level characteristics and human visual perceptions
in CBIR methods. This gap makes it difficult to retrieve an
Mohammed, Oraibi & Hussain
exact picture from a sizable database, a problem that persists
despite the various contributions of existing CBIR algorithms
to image representation and similarity measure. Moreover,
while the Bag of Visual Features (BoVF) model has been
spatial information and lacks semantic meanings. This lack
of spatial and semantic information leads to a less accurate
representation of images, thereby reducing the effectiveness
of the retrieval process. Another model, the Object Bank (OB)
model, provides a high-level picture representation but leads
mensionality can complicate the retrieval process and increase
computational requirements. Lastly, CNN-based Deep Learn-
ing models, despite their effectiveness in scene categorization,
have their own limitations. The complicated training proce-
dure for parameter adjustment, the requirement for enormous
nificant drawbacks of these models. As a result, CNN-based
models cannot be recommended as the best option for CBIR
on various datasets. These problems collectively present a
substantial challenge for the development of efficient and
accurate CBIR systems. In this paper, we contribute to the
field of CBIR by introducing a novel method that leverages
advanced models such as Inception and Xception for feature
extraction from images. Our method addresses the seman-
tic mismatch between an image’s low-level visual qualities
and its high-level semantic content, a significant challenge
in current CBIR algorithms. We provide a comprehensive
analysis of our method’s performance across multiple image
classes, demonstrating its effectiveness and potential for im-
provements in certain areas. The rest of the paper is divided
of the related work of the existing CBIR methods. The third
section, will give a brief overview of what CBIR is and how
it works. The fourth section will go into more detail about
the methods used in this research, including deep learning
techniques, the dataset used, and the approaches taken. The
fifth section will present the results of the research and com-
pare them to other methods. Finally, the conclusion and future
work section will summarize the findings and discuss potential
areas for future research.
II. RELATED WORK
the cutting-edge CBIR methods are critically examined in this
part. A variety of properties, including color, form, texture,
and spatial arrangement, have been incorporated in existing
CBIR algorithms. Similar to this, other interest points-based
features descriptors have been suggested as a method of ob-
taining the attributes for picture retrieval [13, 14] [13, 14]. In
order to recover pictures, scientists in [15] suggested a Micro
Structure Descriptor (MSD) that is generated utilizing edge