978-1-5386-9346-9/19/$31.00 ©2019 IEEE
Content-Based Image Retrieval using Local Patterns and
Supervised Machine Learning Techniques
Maher Alrahhal
1
, Supreethi K.P.
2
1,2
CSE Dept., JNTUH Collage of Engineering, Hyderabad 500085, Telangana, India
1
maherrahal92@gmail.com,
2
supreethi.pujari@gmail.com
Abstract: CBIR is a very important domain, especially in the
last decade due to the increased need for image retrieval
from the multimedia database. In general, we extract low
level (color, texture, and shape) or high-level features (when
we include machine learning techniques) from the images. In
our work, we proposed a new CBIR system using Local
Neighbor Pattern (LNP) with supervised machine learning
techniques. We evaluated the performance of this system by
comparing the system with Local Tetra Pattern (LTrP) using
Corel 1k, Vistex and TDF face databases. We used three
types of the database (i.e color, texture and face databases) to
improve the effectiveness of our system. Performance
analysis shows that LNP gives better performance regarding
the average recall than LBP, LDP, and LTrP. To increase
the accuracy of this system we used the LNP method with
machine learning techniques and performance analysis
shows that local pattern with machine learning techniques
improves the average accuracy from 36.23% to 85.60% when
we use LNP with cubic SVM on DB1 (Corel1K), and from
82.51% to 99.50 % when we use LNP with fine KNN on DB2
(Vistex DB), and from 56.63% to 95% when we use LNP with
ensemble subspace discriminant on DB3 (face DB).
Keywords: Content-Based Image Retrieval, Local Neighbor
Patterns, Supervised Machine Learning Techniques,
Classification, MATLAB.
I. INTRODUCTION
Image Retrieval (IR) means the process of searching the
related images in images Database, and retrieve the most
similarity image to the user. IR techniques can be classified
into Text-Based or Content-Based Image Retrieval. TBIR is
the process of manually adding annotation or description to the
image in the database to describe the content of images and
sometimes for describing other metadata of images like size,
dimensions, and format of images. The advantages of the
TBIR system are simple and very fast in displaying the result
and depend on matching a textual query with description of the
images. However, this method has many disadvantages, such
as the error rate is high and a great amount of manpower and
material resources are needed [1]. Also, adding a description of
the image depend on our points of view, or how we understand
the image, so the description of the image may differ from one
person to another.
Finally, the big issue is adding an annotation to the image
depending on your languages.
CBIR is a method that is used to find an image from a set of
the large database as per the user requirement. CBIR is also
known as query by image content (QBIC) [2]. CBIR ([3], [4],
[5], [6], [7]) is an alternative to the traditional TBIR which
overcomes the above limitations. CBIR covers several areas
like image segmentation, extracts a feature from the image and
converts this feature into a semantic feature. In general, we
extract low level (color, texture, and shape) or high-level
features (when we include machine learning techniques) from
the images [8]. In general, CBIR consists of two phases [9]
off-line phase for feature extraction and online phase for image
retrieval. In the offline phase the system extracts the feature
from all the images in the DB and store them in DB. In on-line
phase, the user inputs a query and the system extracts the
features from this image and measure the similarity by
calculate the distance between query image (i.e feature vector)
and all images in DB (i.e feature vectors), then sort the
distance in ascending way and retrieve the top k images to the
user. Another technique in image retrieval is combining text-
based and content- based to increase accuracy and get better
performance.
CBIR has been used in several fields, such as satellite images
[10], remote sensing, medical imaging[11], fingerprints
scanning ([12], [13]) and biodiversity information systems.
CBIR techniques are being used in the area of satellite images
to find earth minerals, aerial survey, for monitoring
agriculture, to generate weather reports and for tracking
surface objects. Medical imaging is one of the prominent areas
of application of CBIR which can be used for monitoring
patient health reports, to aid diagnosis by identifying the
similar past cases etc. When given a fingerprint query image
CBIR systems can be used to extract the similar fingerprint
images that results in verification of an individual. Fingerprints
are used in banking sector,colleges, corporate companies, and
forensic labs .
II. RELATED WORK
For many years, researchers have been using different local
descriptor methods for image retrieval and classification.
Image classification is supervised machine learning, for that in
training step, we have data sample with a label (class name) for