Page 151 - 2023-Vol19-Issue2

147 |
orientation and color characteristics. This method, however, is
unable to make use of an image’s global properties to exploit
the relationship between the locations of disparate objects. To
protect the privacy of user photos, authors in [16] presented
a CBIR approach for cloud computing-based models. The
visual characteristics were extracted and encoded using k-NN,
and the relevance of the recovered photos to the query image
was calculated using these features. To stop unauthorized
copying of the returned photos, a water marking-based proce-
dure was implemented. However, this water marking technol-
ogy has a weakness in its ability to evaluate in the presence of
distorted geometric elements. Because of its great discrimina-
tive capacity, the Bag of Visual Features (BoVF) model has
been extensively employed in existing CBIR techniques and
has proven to be highly helpful in tasks like object identifica-
tion, automatic picture annotation, and image classification.
These visual feature-based methods have the drawback of ne-
glecting spatial information [17, 18]. Additionally, semantic
meanings are missing from the BoVW model representation.
In order to overcome the problems with spatial and seman-
tic information that BoVW models encountered, the Object
Bank (OB) model was utilized with high level picture repre-
sentation which leads to large dimensionality difficulty when
applied [19–21].
Recent studies [15, 22] show the effectiveness of DL tech-
niques for scene categorization. However, the complicated
training procedure for parameter adjustment, the requirement
for enormous amounts of training data, and excessive training
time are important shortcomings of CNN-based DL models.
As a result, CNN-based models have their own limitations
and cannot be recommended as the best option for CBIR on
various datasets [23,24]. For image retrieval and classification, ated using a variable number of visual words. The results
existing CBIR algorithms have also utilized transform-based
techniques.
Authors in [25] have demonstrated how well CNN extracts
high-level characteristics for picture recall. The generaliza-
tion capacity and performance of these CNN-based models
still need to be improved, though. A two-stage CBIR method
based on EL was recently suggested [24]. The first step in-
volves feature extraction using a CNN-based model, and the
second stage used EL to boost the retrieval system’s efficiency.
The findings demonstrated that, when compared to conven-
tional algorithms, the suggested algorithm still had poorer
picture retrieval and generalization capabilities. The study
did, however, also examine two EL-based CBIR algorithms,
Bagging CNN and Adaboost CNN. Although Bagging CNN
outperformed Adaboost CNN, the total findings were unsat-
isfactory, showing that the suggested algorithm still needs
work. As a result, we suggest a novel CBIR approach in this
study that is based on EL and addresses the shortcomings of
the earlier algorithm. Our suggested method utilizes a more
Mohammed, Oraibi & Hussain
complex EL strategy that adjusts the weights of samples based
on their resemblance to the question picture, as well as an in-
novative mix of CNN-based models and clustering techniques
for feature extraction. We assess our suggested algorithm
using a variety of measures and contrast it with the prior algo-
rithm and other cutting-edge CBIR techniques. The outcomes
demonstrate that our suggested approach works better than
the existing algorithm, achieving higher precision and quicker
retrieval time. Additionally, the stability of our suggested
method to different situations and datasets demonstrates its
strong generalizability.
A privacy-preserving CBIR (PP-CBIR) approach has been
proposed in [26] which offers a valuable solution to the chal-
lenges faced in image retrieval, particularly in terms of privacy
and computational efficiency. This study demonstrates signifi-
cant improvements in both retrieval precision and scalability
while ensuring the protection of sensitive image data. The au-
thors propose an innovative method that represents each image
as a compact aggregated vector derived from local descrip-
tors, effectively reducing computation and communication
costs. The asymmetric Scalar-Product-Preserving Encryption
(ASPE) algorithm is employed to secure these aggregated vec-
tors allowing for similarity computation between encrypted
vectors without the need for decryption or additional com-
munication rounds. This approach effectively addresses the
privacy concerns associated with utilizing cloud servers for
computational tasks. Furthermore, the authors construct a tree
index by recursively clustering all encrypted feature vectors
using the k-means algorithm to enhance search efficiency. The
experiments conducted in the paper utilize two popular local
descriptors, ORB and SIFT, with aggregated vectors gener-
of this study clearly demonstrate the practical value of the
proposed PP-CBIR scheme, offering an effective solution for
securely searching and retrieving image databases in a cipher
text format. The scheme not only maintains privacy but also
improves indexing and retrieval speeds compared to previ-
ous methods. This paper serves as a valuable reference for
the development of privacy-preserving image retrieval meth-
ods, further advancing the field of image processing and data
security.
Authors in [27] proposed a novel approach called the
DTLDN-CBIRA model. This model addresses the need for
effective CBIR techniques specifically designed for plant dis-
ease detection. While existing literature lacks focus on CBIR
for plant diseases, the DTLDN-CBIRA model aims to fill this
gap. In order to overcome the challenge of limited samples
in the dataset, data augmentation techniques such as rotation
and flipping are applied. The DTLDN-CBIRA model utilizes
DenseNet-201 as a feature extractor, taking advantage of its
densely connected network architecture. The hyper parame-