IET Image Processing

Research Article

Comparative analysis of texture feature ISSN 1751-9659

Received on 13th August 2019
Revised 3rd March 2020
extraction techniques for rice grain Accepted on 7th April 2020
E-First on 20th August 2020
classification doi: 10.1049/iet-ipr.2019.1055
www.ietdl.org

Kshetrimayum Robert Singh1* , Saurabh Chaudhury1

1ElectricalEngineering Department, National Institute of Technology Silchar, Silchar 708810, India
*Current affiliation: Department of Electrical Engineering, Mizoram University, Aizawl-796004, India
E-mail: [email protected]

Abstract: Classifications of eight different varieties of rice grain are discussed in this study based on various texture models.
Four local texture feature extraction techniques are proposed and three sets of texture features (SET-A, SET-B and SET-C) are
formed, for the classification task. Performances of the proposed feature sets are compared with the existing techniques based
on, run length matrix, co-occurrence matrix, size zone matrix, neighbourhood grey tone difference matrix and wavelet
decomposition, towards classification of rice grain using a back propagation neural network (BPNN). The proposed techniques
are also tested against publicly available data from Brodatz's texture data set and their results are compared with other
techniques. The classification accuracy by the BPNN classifier is also compared with other statistical classifiers namely, K-
nearest neighbour, linear discriminant classifier and Naive Bayes classifier. It is found that, the proposed feature sets yield better
classification results on both rice data and Brodatz's data. Results show that, feature SET-B, is able to classify rice grain with an
average classification accuracy of 99.63% with a minimum of six features.

1 Introduction image are quite limited and are presented in [35, 36]. In order to
carry out a classification task, one need to look for certain
Visual information plays an important role in understanding the attributes that can best distinguish objects to be classified. It is
surrounding as major portion of human perceived data and understood that texture features will be useful for such applications
information is visual in nature. Broadly speaking, image processing where colour become less discriminative and vice versa. It is found
is an area that deals with improving the quality of pictorial from literature that grey level co-occurrence matrix (GLCM) and
information for better understanding and interpretation. Another grey level run length matrix (GLRLM) are widely used for
objective is to facilitate the automatic machine interpretation of extracting texture features. Applications of GLCM and GLRLM
image [1]. Image understanding applications identify and analyse towards texture feature classification of different types of cereal
the spatial relations among the objects that are present in the grains based on single kernel image are presented in [13, 18, 19,
image, just like the human visual system. An automated non- 23, 24, 29, 37]. Classification of bulk images of cereal grains using
destructive inspection system becomes important, where texture features based on GLCM and GLRLM is presented in [31–
information need to be visually obtained repeatedly and 34, 38, 39]. Rice grain classification based on bulk images using
monotonously [2–4]. Automatic classification system also provides GLCM texture is presented in [36]. Applications of GLCM based
rapid and hygienic inspection [2–6]. Image processing and texture features for classification task other than cereal grains are
computer vision become a promising technology towards also presented in [40–44].
identification, classification and quality assessment for various Most of the cereal grain classification tasks carried out earlier
food products [7–12]. In the case of grain such as rice and wheat, were based on different varieties of grains (namely, wheat, barley,
identification and classification is essential as the end user depends oats, rye, red gram, green gram, ground nut, corn, pearl millet,
on used of a particular variety [13]. Research has been carried out peas, horse gram etc.) having distinguishable colour, shape and
for the past few years towards an intelligent classification system, texture. Classification of grain belonging to a particular species
which can provide solution to prevent human errors during such as rice is rather challenging as their morphologies, colour and
continuous identification and classification process [14]. However, texture are not governed by a specific mathematical function. The
classification task in the case of cereal grain is still challenging, performance of a classifier wholly depends on the nature of their
because of their natural varying appearance [15]. Most of the cereal inputs. If the input data are well separated and distinguishable, then
grain classifications were based on single kernel image, which the classification would yield better results. It is learnt from
requires arranging the grains in a non-touching fashion. literature that same pattern recognition system may not be suitable
Classification of wheat based on single kernel images is carried out for all types of application, thus the processes of pattern
in [14–18]. Classification of different varieties of grains based on recognition and classifications are application specific [45]. This
single grain kernel is also presented in [19–25]. Works related to paper aims to classify eight different varieties of rice grain based
rice grain classification based on single rice kernel are presented in on bulk image. The work also proposes four different rotation
[3, 4, 26–30]. It is difficult to implement such a system because of invariant texture feature extraction schemes and derivation of three
cumbersome setup requirements. Such a system also requires different sets of texture features. The paper compares the
image pre-processing steps like segmentation, background performance of the proposed texture feature sets with that of
removal, noise filtering and object extraction, which is indeed a existing techniques namely, GLCM, GLRLM and other state-of-
time-consuming operation. On the other hand, if the classification the-art techniques. A back propagation neural network (BPNN)
task is carried out using images of bulk samples, many of the with adaptive threshold output is used for the classification task.
requirements of the previously described system become redundant The performance of this classifier is also compared with that of
[31]. Classification and identification of various types of bulk other statistical classifiers namely, K-nearest neighbour classifier
grains are presented in [31–34]. However, it is found that works (KNN), linear discriminant analysis classifier (LDA) and Naive
related to identification and classifications of rice based on bulk Bayes classifier (NB). Unlike most of the classification task

IET Image Process., 2020, Vol. 14 Iss. 11, pp. 2532-2540 2532
© The Institution of Engineering and Technology 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 12,2020 at 10:19:53 UTC from IEEE Xplore. Restrictions apply.
Fig. 1 Different stages involved in the classification task

and the camera is mounted at a height of 7 cm above the target

object. In view of computation time, the acquired colour images
were resized to 200 × 200. Sample images of eight varieties of bulk
rice grain images are presented in Fig. 2. Samples of eight different
standard texture images from Brodatz's texture database are also
presented in Fig. 3.

2.2 Features extraction and classification

Features play significant role in object recognition and
Fig. 2 Sample images of eight different varieties of rice classification. The most important part in object recognition is to
decide the set of features that is helpful in object classification.
Any characteristic or attribute of an object that helps to distinguish
or discriminate an object from other objects is called an image
feature. The purpose of an image feature is to describe the object in
a meaningful manner so as to aid the recognition processes [1].
Feature extraction involves extracting and generating features in
order to aid the classification task. This phase is crucial because the
effectiveness of a classification task depends on the types of the
features. The variation in the intensity and colour profile of an
image give rise to texture of that image, which are nothing but
repeated patterns. This section describes extraction of texture
features using different techniques.
Fig. 3 Texture images from Brodatz's texture database
2.2.1 Grey level co-occurrence matrix: Texture information of
reported in the literature, the performance of the proposed texture an object can be obtained using a GLCM. The matrix represents
feature sets discussed in this work is also tested on publicly the frequency of occurrence of pixel pairs in a particular direction.
available standard images from Brodatz's texture database. In order to reduce the computational time, the input 2D images
were quantised to a grey level equal to 64, so we have a 64 × 64
GLCM. The frequency of occurrence of pixel pairs is taken at 0°,
2 Materials and methods 45°, 90° and 135°, respectively. The average of these four GLCMs
This section presents the process of acquiring colour images of at different angles is evaluated and nine different statistical features
eight different types of rice grain followed by texture features namely, mean, variance, range, energy, contrast, inverse difference
extraction using different techniques, which will be explained in moment, homogeneity and correlation are extracted [32, 44].
the subsequent sections. Finally, the different varieties of rice grain
are classified based on the above-mentioned texture features. The 2.2.2 Grey level run length matrix: GLRLM also provides
performance of these texture features towards rice grain texture information of an image, as it carries information about the
classification are also analysed and compared using standard data length of connected pixels having the same grey level in a specific
set. The proposed classification model is shown in Fig. 1. direction. Similar to GLCM, four GLRLM were generated from the
original 2D image at grey level 64. The average of these four
2.1 Image acquisition GLRLMs is calculated and 11 features namely, short run, long run,
grey level non-uniformity, run length non-uniformity, run
The process involves acquiring video images of eight different percentage, low grey level run, high grey level run, short run low
varieties of rice grains followed by extracting hundred numbers of grey level, short run high grey level, long run low grey level and
frames, each for the eight rice types in MatLab platform. The video long run high grey level were extracted from the resulting GLRLM
images are captured using Coolpix S2500, Nikon digital camera at [46, 47].
12 mega pixels having CCD image sensor size of 28.0735 mm2.
The images are taken outdoor under natural lighting condition.
An adjustable steel camera stand is used for mounting the camera
IET Image Process., 2020, Vol. 14 Iss. 11, pp. 2532-2540 2533
© The Institution of Engineering and Technology 2020

Authorized licensed use limited to: Carleton University. Downloaded on September 12,2020 at 10:19:53 UTC from IEEE Xplore. Restrictions apply.