ISSN 1060-992X, Optical Memory and Neural Networks, 2023, Vol. 32, No. 1, pp. 39–57.

© Allerton Press, Inc., 2023.

Detection and Prediction of Rice Leaf Disease

Using a Hybrid CNN-SVM Model
Devchand J. Chaudharia, * and K. Malathia
a Departmentof Computer Science and Engineering, Saveetha School of Engineering,
Saveetha Institute of Medical and Technical Sciences, Chennai, Tamilnadu, 602117 India
*e-mail: [email protected]
Received May 31, 2022; revised December 12, 2022; accepted December 27, 2022

Abstract—Agriculture is one of India’s greatest money makers and a measure of financial growth. Rice
is one of India’s most widely grown crops as a staple diet. Rice crops have been shown to be heavily
afflicted by illnesses, resulting in significant losses in agriculture. Rice leaf diseases not only cause a
loss of revenue for farmers, but they also decrease the quality of their final output. External appear-
ances of diseased rice leaves can be subjected to image processing processes. On the other hand, dis-
ease sickness may vary depending on the different leaves. Each disease has its own distinct features,
some of leaves have the same colour but various shapes, while others have different colours but the
same shapes. Farmers are sometimes confused and unable to make an accurate judgement when it
comes to pesticide choosing. To solve this problem, a hybrid CNN (Inception-ResNet)-SVM model
for detecting and treating damaged rice leaves has been developed. In this designed model, the images
are collected and gathered by capturing rice leaf using camera at the agricultural field. These images
are refined to improve image quality and visibility for reliable estimation, and then segregated using
Grab-Cut algorithm to eliminate undesired sections of image. Features of the segmented images are
extracted and classified using hybrid CNN (Inception-Resnet V2)-SVM algorithm. The developed
model’s study results are analysed and discussed to recent techniques. The suggested model achieved
accuracy, precision, recall, and error values of 0.97, 0.93 and 0.03 accordingly. As a conclusion, sug-
gested model outperforms revious methodologies.

Keywords: rice leaf disease, contrast limited adaptive histogram equalization, wiener filter, CNN-
SVM, grab-cut algorithm
DOI: 10.3103/S1060992X2301006X

1. INTRODUCTION
India is a rapidly developing country with agriculture serving as the country’s backbone. However,
there are other difficulties in the agricultural area, including significant productivity losses [1]. Rice dis-
eases have substantial negative effect on quality and amount of rice crops, and are major threat to food
security. In contemporary agriculture, collecting details concerning real-time rice health and growth is
prioritized [2]. Rice illnesses can be predicted before they emerge throughout the rice production process
using data from several sources [3]. Visual findings by skilled farmers or rice specialists on field used to
substantially in conventional approach of detecting rice diseases. This necessitates continual skilled mon-
itoring, that prohibitively expensive in big farms [4]. Farmers in certain emerging regions have to travel
long way to get agronomic specialists, which is moment, costly, and uncertain.
A new technique for detecting and treating plant diseases has developed as result of rapid advancement
of image processing and pattern identification systems [5]. Image recognition utilizing deep learning is an
essential research topic because it may aid monitor massive areas of plants and identify diseases on plant
leaves manually [6]. Several previous studies conducted to evaluate image identification utilizing a specific
classifier that distinguishes between image types [7]. Due to improvements in digital cameras and risen
computation power [8] auto image processing innovation shown strong performance in current decades
and successfully deployed in a number of fields, such as medical image analysis, food interpretation, bio-
metrics, factory output, and expression learning, and other.
Most plant illnesses may be identified automatically using effective image recognition techniques and
the primary source of detection is usually the plant leaves [9]. The rice leaf plant diseases Leaf smut, Rice
burst, Bacterial blight, and brown spot are identified utilising hybrid CNN (Inception-Resnet v2)-SVM

39
40 CHAUDHARI, MALATHI

with actual pictures from farm areas in this suggested model. Every disease has own distinct features; for
example, some diseases have similar colour but various shapes, while others have different colours but
similar shapes [10]. So, to classify the disease accurately from the captured images some of the pre-pro-
cessing techniques are used to enhance the image. The enhanced images consist of unwanted portions of
the leaf that should be segmented to avoid wrong predictions and low computational time. Major contri-
butions of the proposed model are:
• Detection and prediction of rice leaf plant disease using Hybrid CNN-SVM model.
• Raw data captured from rice field are pre-processed using image resizing, Wiener filter and image
quality increase using Contrast Limited Adaptive histogram equalization (CLAHE) approaches.
•The GCA is used to divide pre-processed photos and eliminate undesired regions.
• To obtain features from image for categorization, a hybrid CNN depending on type of Inception-
Resnet V2 technique is utilised.
• Extracted features from the hybrid CNN algorithm is classified using the SVM algorithm.
Section 2 describes particular research reports relevant to current solutions utilised for rice leaf disease
detection. Section 3 provides a brief overview of suggested methodology. The suggested framework’s out-
comes and performance measures are detailed in Section 4. The entire study project is wrapped up in Sec-
tion 5.

2. LITERATURE SURVEY
Many different methods are designed for predicting the leaf disease with various techniques and clas-
sifiers. Some of the existing research papers are reviewed below.
Jiang F. et al. [11] had performed deep learning and support vector machines were used to identify leaf
diseases in pictures. The rice leaf disease images are first extracted using convolution neural networks
(CNNs) in this model. After that, the SVM algorithm is used to classify and predict the disease. The opti-
mal SVM model values are identified utilising 10-fold cross validation method. The test results reveal that
deep learning and SVM-based rice disease recognition model has higher average proper recognition rate.
Shrivastava V. et al. [12] had designed rice plant illness categorization using deep CNN transfer learn-
ing especially for Oryza Sativa disease. Photos of sick symptoms in leaves and stems were acquired on rice
field. In field, 619 images of rice plant injury were collected and grouped into four categories: Healthy
Leaves (HL), Sheat Blight (SB), Rice Blast (RB) and Bacterial Leaf Blight (BLB). A pre-trained deep
CNN model is utilized as feature extractor, while SVM is used for categorization. This technique of
advanced rice disease detection used both prophylactic and an early warning system.
Atole R.R. and Park D. [13] had performed for detection of frequent rice plant irregularities, a multi-
class deep CNN classifier was used. The use of deep CNN in categorization of rice plants based on images
of their leaves for their health status is examined. The training employed 600 photos of rice plants to rep-
resent the different classes. Agricultural technicians confirmed training and validation dataset photos were
acquired from rice fields all over district. A three-class classifier depicting normal, diseased, and snail-
infested plants is created utilizing transfer learning from an AlexNet deep network.
Krishnamoorthy N. et al. [14] had designed deep neural networks with learning algorithms were used
to forecast rice leaf diseases. Rice is a key grain crop that provides energy to half of world’s population as
staple meal. Soil fertility, precipitation, pests, temperature, virus bacteria, and other abiotic and biotic fac-
tors impact rice grain harvest production amount and quality. The advancement of agriculture technical
support has made much easier to automatically detect harmful organisms in rice plant leaves. Inception-
ResNetV2 is sort of CNN model used to recognise illnesses in rice leaf pictures using transfer learning
technique.
Larijani M.R. et al. [15] had developed KNN algorithm enhancement utilizing k-means used to exam-
ine an image analysis strategy for detecting rice blast disease in field situations. In order to define disease
patches on rice leaves, colour images made utilizing image processing method and K-means-improved
KNN algorithm used to divide images in Lab colour space. The Otsu approach used to produce an adap-
tive thresholding histogram of photos based on shape or to reduce grey level in binary images. Squared
classification depends on Euclidean distance. Finally specificity, total accuracy and sensitivity of sug-
gested algorithm are evaluated to determine its efficacy.
Most of existing techniques for designing rice leaf plant disease detection are done by using CNN-SVM,
CNN (AlexNet)-SVM, CNN (Alexnet), CNN (Inception-ResNet) and KNN. From the mentioned
existing techniques, the proposed CNN (Inception-ResNet)-SVM model performs better.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 41

Data
collection

Pre-processing

CLAHE (Contrast Limited

Image Resizing Weiner Filter Adaptive Histogram
Equalization

Segmentation

Grab cut algorithm

Classification

Hybrid CNN -SVM

Predicted output

Bacterial
Rice Blast Leaf Smut Brown Spot
Blight

Fig. 1. Suggested model’s process flow.

3. PROPOSED METHODOLOGY
Rice is an essential grain since it feeds humans with energy. The identification and prediction of Rice
leaf diseases in farming fields is an important process for appropriate management and reduction of Rice
damage. Rice leaf infections can impact leaf at any time during planting and harvesting, resulting in sig-
nificant losses in rice farming and market price. Rice leaf diseases not only waste farmer’s income, but also
they reduce quality of their final product. Hence, Image processing operations can be applied on external
appearances of infected rice leaf to detect and predict disease at early stage.
Figure 1 depicts suggested model’s process flow. In this designed model, the data are collected by cap-
turing images of the rice leaf using camera at the agricultural field. These collected images are pre-pro-
cessed using the image resizing, wiener filter and CLAHE techniques. The pre-processed images are seg-
mented using the Grab-Cut algorithm for segmenting the rice leaf form unwanted portions of the image.
Features of the segmented images are extracted and classified using hybrid CNN (Inception-Resnet V2)-
SVM algorithm.

3.1. Data Collection

Collection of data is the process of gathering the necessary data, such as text or images that must be
predicted. The data from the camera is collected in this model to detect face expressions and gender iden-
tification.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

42 CHAUDHARI, MALATHI

3.2. Pre-Processing
Data collection consists of raw data that cannot be classified with high accuracy by a classifier. In order
to get highest precision, specific pre-processing procedures are used.
3.2.1. Image resizing. Process of resizing source input image known to as image resizing. The original
input image is scaled into the size of 256 × 256 pixels [16]. This type of picture size makes deep learning
techniques like CNN to employ easily. Bi-cubic interpolation produces a sharper edge result than prior
methods such as bilinear interpolation. The mathematical term used to represent image resizing is given
below (1).
3 3
p ( A, B ) = x A B .
i j
ij (1)
i =0 j =0

From Eq. (1) B signifies the height in pixel, A denotes the width in pixel and xij denotes the squared
area in the image. The noise is removed using a filtering algorithm using the scaled image as an input.
3.2.2. Wiener filter. The Wiener filter is low-pass filter used to recover noise-degraded signals in variety
of situations. The filter defined on statistical approach where signal and noise are considered to stationary
linear random processes with specified spectral properties [17]. This approach outperforms others because
it achieves best balance in bias–variance trade-off. In terms, Wiener filter is an adaptive filter that evalu-
ates neighbourhood’s mean before providing little flattening when variance is large and greater flattening
when variance is low.
The filter reduces difference between evaluated and actual signals much as feasible. The error ratio for
untainted image f and anticipated image f is as regards:

e2 = E {( f − f ) }. 2
(2)

Where E{∙} denotes arguments anticipated value. Obtaining lowest of quadratic error function becomes
difficulty of getting an approximated image. The frequency domain utilised to do this, using following
indicators: noise and image are nonstationary, noise or picture has a mean of 0, and intensity levels in pre-
dicted image reduced by linear function. The error functions minimal is provided while following factors
closely:
 H * (u, v ) v S f (u, v ) 
Fˆ (u, v ) =   G (u, v ) . (3)
S f (u, v ) H (u, v ) + S f (u, v )
2

Which Fˆ (u, v ) is predicted picture in spectral domain, H(u, v) is deterioration function’s transform, G(u, v)
is reduced picture’s transform, H * (u, v ) is complex conjugate of H(u, v), and S f (u, v ) = |F(u, v)|2 is non-
degraded picture’s power distribution. The combination of complicated value and its inverse is magnitude
of complicated number doubled, as per general principle of filter. Thus,
 H (u, v ) 
2
Fˆ (u, v ) =  1  G (u, v ) , (4)
H (u, v ) H (u, v ) + Sη (u, v ) S f (u, v )
2

where the noise power spectrum is Sη (u, v ) = N (u, v ) . Since spectrogram of non-degraded image is
2

infrequently known, the expression S η (u, v ) S f (u, v ) is supplanted among constant K.

The Wiener filter connected to images deteriorated by stable power cumulative noise in digital image
processing. The diameter of the neighbourhood and power of noisy are the variables of Wiener filter.
3.2.3. CLAHE (Contrast Limited Adaptive Histogram Equalization). Every histogram of sub picture
disperses brightness values of pixels in adaptive histogram equalization, which enhances contrast level and
shows more features [18]. When applied on chaotic pictures, like medical images, this technique may
increase noise. CLAHE is an expanded form of adaptive histogram equalization. The CLAHE algorithm,
in reality, spreads visual quality and limiting noise enhancement. CLAHE improves distinction between
background of an image and objects. It limits hurdle of noise and contrast intensification as well prevents
problem of oversaturation in similar regions. The number of sub pictures and Clip Limit have significant
impact on CLAHE method’s performance. The pictures are separated into number of sub-images (tiles),
and contrast transform function is calculated for every pixel using 'Clip Limit’ intensity factor. The Expo-

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 43

nential, Uniform, or Rayleigh distributions are used to create contrast transform function. The CLAHE
method comprises several steps:
Its first phase to partition actual image across M × N tiles and compute histogram for every zone based
on image grey levels. Then, for each region, compute contrast-limited histogram:
N avg = ( N x × NY ) N grey . (5)
To compute mean value of pixels, N grey by number of grey levels in region and umber of pixels within
X and Y axes.
N avrgrey = N  clip N grey . (6)
NCL represents actual clip limit CL, while N clip is standardized CL in [0, 1] range. If pixel value is more
than NCL , pixels are clipped. The median of remaining pixels and overall number of clipped pixels N  clip .
Hregionclip (i ) = Hregion (i ) + NCL. (7)
The clipped histogram is Hregionclip (i ) and original histogram is H region (i ) of every region at ith grey level.
Step= N grey N remain . (8)
Step is positive integer higher than or equal to 1, where N remain is number of clipped pixels left. Using
preceding step, proceed from lowest to highest grey level. Allocate one image to grey level if number of
pixels in grey level is less than N CL . To compute new path and create a new search session when all pixels
will not dispersed at end of search. When all of remaining pixels were distributed, stop. Using transform,
increase intensity values in each zone (e.g., Rayleigh).The Rayleigh transform is provided by:
 
y (i ) = ymin + 2α ln  1
2
, (9)
1 − pinput (i ) 
( y (i ) − ymin )
 ( y (i ) − ymin )2 
p ( y (i )) = exp  −  for y (i ) ≥ ymin . (10)
α2  2α2 
Transform clipped histogram into a continuous possibility pinput (i ) to generate a transfer function.
Assuming, α is Rayleigh scaling parameter and ymin is bottom bound of pixel value.
The linear contrast extension used to eliminate suddenly shifting effects.
x (i ) − xmin
y (i ) = . (11)
xmax − xmin
Reduce border distortions by using bilinear interpolation to determine grey level attribution of pixels.

3.3. Segmentation
Segmentation is mainly useful for eliminating the unwanted portions of the image for reducing the
computational time and wrong predictions. The GAC used for classification in this model.
3.3.1. Grab Cut algorithm. Grab Cut is method for image separation that is repetitive and automated.
To create perfect image segmentation, that graph is constructed by employing a small cost reduction func-
tion. The image pixels are used to build graph vertices. In other terms, graph node represents each pixel in
image. Two additional nodes, sink and source, are inserted to graph in contrast to such nodes. Every pixel
in image, or node in graph, is linked to these two nodes. The foreground connection point is defined by
source node, while background connection point is represented by sink node. The graph’s edge weights
are generated using cost function based on region and limit details in image. The graph segmented utiliz-
ing Min-Cut/Max-Flow method. The Grab Cut approach utilizes Gaussian Mixture Models (GMMs) to
determine area information from an image’s colour information. The algorithm’s mathematical statement
is as follows. I, p = (p1, p2, ..., pN) of N pixels where pi = (Ri, Gi, Bi), i ∈ [1, ..., N] in the RGB colour space.
The categorization is represented as an array s = (S1, S2, ..., SN), Si ∈ {0, 1} with label provided to each
pixel of image expressing its connection to background or foreground. The user starts procedure by semi-
automatically creating rectangle (R) contains area to be divided. The image is separated into three sections
termed RB, RF, and RU, which reflect initial values of background, foreground, and unknown pixels,
respectively, once user specifies R. The images on outside of R is designated RB, while images inside are

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

44 CHAUDHARI, MALATHI

marked RU. Grab Cut determines RU images are in foreground or background. It uses colour data pre-
sented by GMMs for this goal. For foreground pixels (si = 1), complete covariance GMMs of C parts is
generated, and for background pixels (si = 0), limited as follows:
θ = {π ( s, c ) , ω ( s, c ) ,  ( s, c ) , s ∈ {0,1} , c = 1,.., C} . (12)
Where ω denotes GMMs’ means, P represents model’s covariance matrices and π denotes weight of
model. The array c = (c1, ..., ci, ... cN), ci ∈ (1, ..., C), i ∈ [1, ..., N] indicates component of foreground or
background GMMs to which pixel pi belongs. For segmentation, the energy function revealed is:
E ( s, c, θ, p) = U ( s, c, θ, p) + V ( s, p) . (13)
Based on probability distributions pr (.) of GMMs, U is likelihood capacity.
U ( s, c, θ, p ) =  − log pr ( p
i
i si , ci , θ ) − log π ( si , ci ) . (14)

Where V is regularising prior that predicts segmented regions to scolour consistent, with neighbourhood
C about each pixel.
V ( s, p) = γ  [Sn ≠ Sm ] exp ( −β pm − pn
2
). (15)
{ }
m,n ∈C

The picture with specific base rectangle is subjected to energy minimization approach.

3.4. Feature Extraction

The features are obtained from pre-processing results in this procedure. The feature extraction process
is done by using modified CNN (Inception-Resnet V2).
3.4.1. CNN (Inception-Resnet V2). The earliest layers of Inception-ResNet v2 model are intended to
gather low-level characteristics such as dots, lines, and edges [20]. The network’s higher layers also will
gather middle-level attributes in particular areas of image, like sharpness, texture, and picture shadowing.
Finally, higher layer will acquire high-level information from retina image, like shape, as diagnose sick-
ness. The operations involved Inception-ResNet v2 based model has been explained in the following sub-
sections.
3.4.2. Inception. In the design of CNN, feature extraction is done using inception structural designs.
The variation lies in the part of feature extraction part. Inception v2 utilizes base layer and filter sequence
to extract information. The Inception model’s purpose is to increase resource removal in system. When
compared to shallower and less wide networks, fundamental advantage of Inception layout is that offers
considerable value rise at variable gain in computing needs. By employing effective factorization tech-
niques, inception design tries to reduce the limitation and increase the computation complexity. To
improve calculation speed, 5 × 5 image convolution layer in Inception v2 model is parameterized for 3 ×
3 image convolution.
3.4.3. Factorizing convolutions with large filter size. The various methods of factorising convolutions in
different contexts have been defined in this section, with the purpose of improving the software’s comput-
ing performance. Due to the complete convolutional nature of the Inception model, each weight reflects
a multiplication per activation. As a result, any reduction in computation cost results in a reduction of vari-
able count. It states that along with appropriate factorization, more extricated variables are obtained
resulting in greater training rate. Furthermore, the computation and memory savings can lead to a larger
filter bank while allowing each model copy to be trained on its own machine.
3.4.4. Factorization into smaller convolutions. In concerns of computation, convolutions with large spa-
tial filters (e.g. 5 × 5 or 7 × 7) tended to be unfairly expensive. A 5 × 5 convolution with n extracts over
grid via m extracts, for example, is 25/9 = 2.78 times more computationally expensive than 3 × 3 problem
with same set of extracts. Obviously, 5 × 5 filter can collect requires among impulses among constructions
of units further away in previous layers, thus a reduction in arithmetical size of filters seems to have great
rate of expressiveness. However, it may test if a 5 × 5 convolution could returned with multilayer network
utilizing minus variables because input size and depth are same. When expand into 5 × 5 convolution
computation graph, it observe that each outcome seems as small entirely connected network moving
across 5 × 5 tiles above its attempt, as seen in Fig. 2.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 45

Fig. 2. Mini network replacing 5 × 5 convolution.

3.4.5. Spatial factorization into asymmetric convolutions. Because it simplified to succession of 3 × 3

convolutional layers, 3 × 3 convolution isn’t necessarily beneficial. Instead, it can be factored into 2 × 2
convolutions. Furthermore, it rotates out, which can still be improved by 2 × 2 using asymmetric convo-
lutions, e.g. n × 1. For example, sliding a 2-layer network by same available field in a 3 × 3 convolution
equates to 3 × 1 convolution succeeded by a 13 convolution. The two-layer approaches are 33% lower for
same set of output filters, when number of inlet and outlet extraction are similar. Factorizing a 3 × 3 con-
volution into two 2 × 2 saves just 11% of computing time as shown in Fig. 3.
3.4.6. Efficient grid size reduction. Convolution networks are commonly utilised for numerous pooling
processes in order to lower the grid size of characteristic maps. The activation dimension of the networks
is enhanced to get rid of a representative bottleneck before using maximum or average pooling. For exam-
ple, starting from the l × l grid by k filters, to arrive at a l × l grid with 2k filters, it is needed to calculate
2 2
a stride-1 convolutional through 2k filters and after that employ a further pooling stage. This indicates
which the overall computational prices are controlled with the costly convolutional on the big grid with
2l 2k 2 processes. One opportunity will be to control to pooling through convolution and then resultant in

()
2
2 l k 2 falling the computation rate with a quarter. However, overall dimensionality of demonstration
2

()
2
reduces to l k minus communicative network, this produces representative bottlenecks. Instead of
2
using this, another option is suggested that decreases the calculation rate even more by reducing the rep-
resentative bottleneck.
3.4.7. Inception-v2. Now, the above-mentioned concepts are integrated to present a new structural
design as shown in figure 4. It has three traditional inception sections 3535 by 288 filters in network’s
Inception section. Using grid reduction approach described in Section 2.3.3, it reduced to 17 × 17 grid
with 768 filters. The grid reduction method can reduce it to an 8 × 8 × 1280 grid. Two Inception compo-
nents are present at coarse 8 × 8 level, due to concatenated result filter bank size of 2048 for each tile. It
is observed that the network quality is relatively stable.
3.4.8. Inception-ResNet v2. In comparison to classical Inception network, residual version of Incep-
tion network uses a simpler Inception block. Each Inception phase is followed by a filter-expansion layer
that is used to enhance depth of filter bank before it used to map depthless of input. The dimensionality
minimization achieved by Inception block must be returned. Figure 5 depicts structure of Inception-Res-
Net v2 model.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

46 CHAUDHARI, MALATHI

Fig. 3. Mini Network replacing 3 × 3 convolution.

Batch-normalization is utilized exclusively on top of traditional layers in Inception-ResNet, but not on

top of summations, which is small difference between residual and non-residual Inception variations.
Although it is reasonable to believe that comprehensive batch normalization is desirable, each model
replica must train under its own graphical processing unit (GPU). The number of inception blocks can be
enhanced significantly by avoiding batch-normalization. Convolutional layer, activation layer, and pool-
ing layer are three core layers of Inception ResNet v2.
3.4.9. Convolutional layer. The convolution procedure of input image is used to extract features. The
formula is as follows:
 
x = f (
 i∈Mj
)
X ijnYi n −1 + a nj  .

(16)

Where Y is feature vectors with kernel value X, n denotes number of network layers, bias value a and Mj is
set of each layer feature maps and ∗ denotes convolution operation.
• The input image’s dimensions are (H × W × D).
• ( Fh × Fw × D) is the kernel dimension.
• (( H − Fh + 1) × (W − Fw + 1) × 1) is output dimension of after convolution.
Resize the supplied images to 229 × 229 pixels utilizing three colour channels. The basic goal of using
neural networks is to lower image’s width and height while increasing its depth (store all relevant informa-
tion in depth of image). The depth of image is affected by number of convolution layers applied in each
sector; thus, network contains 825 convolution layers for extracting features.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 47

Filter Concat

3×3
stride 2

3×3 3×3
stride 1 stride 2

Pool stride
1×1 1×1 2

Base

17 × 17 × 640

Concat
17 × 17 × 320 17 × 17 × 320

Conv Pool
35 × 35 × 320

Fig. 4. Inception V2 architecture.

(i) Activation layer

The activation layer stands between (or end of) neural networks, determining whether or not neuron
will fire. The activation layer employs a variety of activation functions. The suggested work’s activation
function is rectified linear unit (ReLu), which means it will only fire on positive values.
(ii) Pooling layer
The pooling layer’s major job is to keep previous features and removing ones that aren’t needed. They
employed network for extracting features and categorization of input data with help of fully connected
layer and softmax in a project based on Inception ResNet v2. Therefore, in proposed study, features are
extracted from average pool layer output and fed to classifiers for categorization. As a result, collected fea-
tures are sent into classifier, which uses to identify retina disease.

3.5. Classification
Classification is technique for predicting class of supplied data points. Classes described utilizing
phrases such as targets, labels, and categories. Classification predictive modelling (y) is process of estimat-
ing a mapping function (f) from discrete input variables (X) to discrete output variables.
3.5.1. SVM Classifier. SVM is machine learning method for identifying data and recognising patterns
in categorise it [21]. It includes assigning set of input data to one of two classes. A training set required to
learn support vector machines, with each element of set specifying which class it belongs to. SVM model,
data from training set is divided by border with broadest possible margin, i.e. the distance from this hyper-

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

48 CHAUDHARI, MALATHI

Filter concat 35 × 35 × 384

3 × 3 conv Maxpoll
(192 V) (stride = 2V)

Filter concat 71 × 71 × 192

3 × 3 conv
(96 V)

3 × 3 conv
(96 V) 1 × 7 conv
(64)

1 × 1 conv
7 × 1 conv
(64)
(64)

l × l conv
(64)

Filter concat 73 × 73 × 160

3 × 3 Maxpool 3 × 3 conv
(stride 2 V) (96 stride 2 V)

3 × 3 conv
(64)
147 × 147 × 64

3 × 3 conv
(32 V)
147 × 147 × 32

3 × 3 conv
(32 stride 2 V)
149 × 149 × 32

Input
(299 × 299 × 3)
299 × 299 × 3

Fig. 5. Inception-Resnet V2 architecture.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 49

plane. The Support Vector Machine (SVM) is well-known binary classification learning technique. The
main purpose is to find the best hyper plane for splitting data into two categories. Combining multiple
binary SVMs to achieve multiclass classification was recently achieved. The basic architecture of SVM
classifier is given as Fig. 6.
Let x is considered as input, j be the training instances {x j , y j } , j = 1,… m and for each occurrence
consist x j as well as label yi ∈ {−1,1} . A weight vector (w), a bias (b) both are used to parameterized every
hyper plane which is expressed as Eq. (4) as given as below.
w.x + b = 0. (17)
The hyper plane function that divides training and testing data can be described by following Eq. (5),
f ( x ) = sgn ( w.x + b ) . (18)
When interacting with kernel functions, prior function can be expressed as an Eq. (6),

 N

f ( x ) = sgn 

α y k ( x , x ) + b  .
j =1
j j j (19)

Where x j is the input of training instance, yi is its corresponding class label, b is a bias, K ( x j , x ) is
employed kernel function that maps input vectors into an extended features space, and N is number of
training instances. The variables α j are determined by two constraints (7) and (8).

0 ≤α j , j = 1,.., N , (20)
N

α y
j =1
j j = 0. (21)

Thus, following equations are utilised to train and test disease prediction feature retrieved from rice leaf
dataset. The algorithm 1 contains pseudocode for predicting rice leaf disease.

Algorithm 1. Pseudocode for rice leaf disease prediction

Input: Dataset = X
#Pre-processing
IR = Image Resizing (X) // By using Eq. (1)
W = Wiener Filter (IR) // By using Eq. (2), (3) and (4)
C = CLAHE (W) // By using Eq. (10)
#Segmentation
GC = Grab cut(C) // By using Eq. (13)
#Feature extraction
FE = Inception-Resnet V2 (GC) // By using Eq. (16)
#Classification
S = Training_SVM (FE)
S1 = Testing_SVM (S) // Prediction of classes by using Eq. (18)
If (S1 = 0)
Class 1 = Blast
Elif(S1 = 1)
Class 2 = Blight
Elif(S1 = 2)
Class 3 = Brown spot
Else
Class 4 = Leaf smut
end
Output: Predicted leaf disease classes

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

50 CHAUDHARI, MALATHI

Bias b
L(X, X1)
J1

L(X, X2) J2
Fully connected layer
J3 6 Output (Y)
(X)
L(X, X3) J4

L(X, Xm) Weights (Lagrange's

multipliers)
Hidden Nodes

Fig. 6. Architecture of SVM classifier.

4. RESULT AND DISCUSSION

Rice leaf dataset is used to validate function and performance of suggested method for rice leaf disease
detection on Python software. The 64-bit operating system, NVIDIA GTX 1650 4Gb (GDDR6) GPU,
CPU @ 2.50 GHz, 16.0 GB Memory (RAM), and Intel(R) Core(TM) i5-10300H processor are used in
testing. In this designed model, the data are collected by capturing images of the rice leaf using camera at
the agricultural field. These collected images are pre-processed using the image resizing, wiener filter and
CLAHE techniques. The pre-processed images are segmented using the Grab-Cut algorithm for seg-
menting the rice leaf form unwanted portions of the image. Features of the segmented images are extracted
and classified using hybrid CNN (Inception-Resnet V2)-SVM algorithm.

4.1. Dataset
For rice leaf disease prediction, data set used in this model was acquired from [22]. This dataset
includes photos for diseases like blast, leaf smut, brown spot, and blight, and also images for testing and
training.

4.2. Pre-Processing and Segmentation

Conversion of sample images using pre-processing techniques such as image resizing, wiener filter and
CLAHE for enhancing the clarity of the image and the pre-processed image is segmented with grab cut
algorithm, which is illustrated in Fig. 7.
The proposed method’s ROC curve is presented in Fig. 8. Area Under the Curve (AUC) or Receiver
Operating Characteristics (ROC) curves appear to be performance metric for classification issues at var-
ious threshold levels. The AUC is measure of separation, whereas ROC is probability curve. The system’s
performance measure is regarded well when curve approaches “1”. When curve’s value is “0”, system’s
performance is examined poor. The ROC value in proposed method is “1”, indicating that system per-
forms well.
Figure 9 depicts suggested model’s confusion metrics. The performance of classification algorithm can
be represented using confusion matrix. Making a confusion matrix can figure out which components of
your classification model are correct and which are wrong. A confusion matrix is a valuable method for
examining classification models. It gives a clear picture of how well the model identified the classes based
on the data supplied it, as well as how the classes were misclassified. To compare values such as True Pos-
itives, True Negatives, False Positives, and False Negatives, confusion matrices are necessary. The cor-
rectly predicted data for class 0, 1, 2 and 3 are 75, 79, 36 and 37.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 51

Original image Image resizing Wiener filter CLAHE Grab-cut

0 0 0 0 0
250
50 50 50 50
500
750 100 100 100 100
1000
150 150 150 150
1250
1500 200 200 200 200
1750
250 250 250 250
0 500 1000 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
0 0 0 0 0

50 50 50 50
250
100 100 100 100
500
150 150 150 150
750 200 200 200 200

0 500 10001500200025003000 250 250 250 250

0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
0 0 0 0 0

50 50 50 50
50
100 100 100 100
100 150 150 150 150

150 200 200 200 200

0 100 200 300 400 500 250 250 250 250

0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
0 0 0 0 0

50 50 50 50
250
100 100 100 100
500
150 150 150 150
750 200 200 200 200

0 500 10001500200025003000 250 250 250 250

0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
0 0 0 0
0
50 50 50 50

50 100 100 100 100

150 150 150 150

100
200 200 200 200
0 100 200 300 400 250 250 250 250
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250

Fig. 7. Conversion of sample images using pre-processing and segmentation.

Receiver operating characteristic

1.0

0.8
True positive rate

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1.0

False positive rate

Fig. 8. ROC curve of the proposed method.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

52 CHAUDHARI, MALATHI

0 75 5 0 0

1 1 79 0 0

True label
2 0 0 36 4

3 0 0 3 37

0 1 2 3
Predicted label

Fig. 9. Confusion metrics.

1.0

0.8

0.6
Aceuracy

0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 10. Comparison of Accuracy for suggested and existing method.

4.3. Performance Metrics

The existing methods compared with proposed method are VGG16, Mobilenet V1, ResNet and Alex-
Net. When compared to existing models, suggested model performs better in terms of recall, sensitivity,
F1 score, error, precision, false positive rate (FPR), false negative rate (FNR), specificity and accuracy.
Figure 10 depicts reliability comparison metrics for proposed and existing techniques. In this, existing
method such as VGG16, Mobilenet v1, ResNet and Alexnet are compared with proposed Inception-Res-
Net V2 + SVM model. The accuracy value for Inception-ResNet v2 + SVM, VGG16, Mobilenet v1, Res-
Net and Alexnet is 0.97, 0.89, 0.87, 0.82 and 0.74. In comparison to existing approaches, suggested model

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 53

0.8

0.6

Precision
0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 11. Comparison of Precision for suggested and existing method.

0.8

0.6
Recall

0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 12. Comparison of Recall for suggested and existing method.

has higher accuracy. Figure 11 explains the comparison of precision metrics, in this the precision value for
(Inception-ResNet v2) + SVM is 0.93. Which is greater compared to VGG16, Mobilenet v1, ResNet and
Alexnet whose values are 0.90, 0.89, 0.83 and 0.7.
The recall metrics for suggested and existing approaches are compared in Fig. 12. Recall metrics for
Inception-ResNet v2 + SVM, VGG16, Mobilenet V1, ResNet and Alexnet 0.93, 0.84, 0.70, 0.71 and 0.68.
When compared to existing methodologies, suggested method has higher recall value. Figure 13 explains
the comparison of error metrics, in this the acquired error metrics for Inception-ResNet v2 + SVM is 0.03,
which is smaller compared to VGG16, Mobilenet V1, ResNet and Alexnet whose values are 0.11, 0.13, 0.18
and 0.26.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

54 CHAUDHARI, MALATHI

0.25

0.20

0.15

Error
0.10

0.05

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 13. Comparison of Error for suggested and existing method.

1.0

0.8
Specificity

0.6

0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 14. Comparison of specificity for suggested and existing method.

The comparison parameters for proposed and existing techniques are explained in Fig. 14. Specificity
metrics for Inception-ResNet v2 + SVM, VGG16, Mobilenet V1, ResNet and Alexnet 0.98, 0.93, 0.83,
0.86 and 0.80. In comparison to existing approaches, specificity value of suggested method is higher. Fig-
ure 15 explains the comparison of F1_score metrics, in this the acquired error metrics for Inception-Res-
Net v2 + SVM is 0.93, which is greater compared to VGG16, Mobilenet V1, ResNet and Alexnet whose
values are 0.90, 0.84, 0.77 and 0.60.
The NPV comparison measures for proposed and existing approaches are described in Fig. 16. The
NPV metrics for Inception-ResNet v2 + SVM, VGG16, Mobilenet V1, ResNet and Alexnet is 0.98, 0.93,
0.83, 0.90 and 0.82. When compared to existing methodologies, proposed method has higher NPV. Figure 17

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 55

0.8

0.6

F1_Score
0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 15. Comparison of F1_Score for proposed and existing method.

1.0

0.8
Negative predictive value

0.6

0.4

0.2

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM

Fig. 16. Comparison of NPV for proposed and existing method.

explains the comparison of False Negative Rate (FNR) metrics, in this the acquired FNR metrics of
Inception-ResNet v2 + SVM is 0.06, which is smaller compared to VGG16, Mobilenet V1, ResNet and
Alexnet, whose values are 0.10, 0.14, 0.28 and 0.63.
Figure 18 depicts False Positive Rate (FPR) comparative metrics for proposed and existing
approaches. The FPR metrics for Inception-ResNet v2 + SVM, VGG16, Mobilenet V1, ResNet and
Alexnet is 0.019, 0.025, 0.038, 0.130 and 0.192. In comparison to existing methodologies, proposed
method has lower FPR value. The suggested model’s performance measures are examined and compared
to existing methodologies.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

56 CHAUDHARI, MALATHI

0.6

0.5

Fslse negatyve rate

0.4

0.3

0.2

0.1

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM
Fig. 17. Comparison of FNR for proposed and existing method.

0.200

0.175

0.150
Fslse positive rate

0.125

0.100

0.075

0.050

0.025

0
(Inception VGG16 MobileNetv1 ResNet AlexNet
ResNetv2)
+SVM
Fig. 18. Comparison of FPR for proposed and existing method.

5. CONCLUSIONS
Fast and precise disease prediction is crucial in agricultural disease management. In this designed
model, the rice leaf images are captured by camera placed in the agricultural filed, these images are col-
lected and gathered as a dataset. The collected images are pre-processed for enhancing the accuracy of the
proposed model. These pre-processed images are segmented using the grab-cut algorithm for removing
the unwanted portions of the rice leaf. The features of the segmented portions are extracted with the CNN
(Inception-ResNet V2) and SVM is used for classifying the rice leaf disease. The suggested model’s per-
formance measures are examined and compared to existing methodologies. The accuracy, precision,
recall, error, specificity, F1-Score, NPV, FNR, and FPR obtained values of 0.97, 0.93, 0.03, 0.98, 0.93,
0.98, 0.068, and 0.019, respectively. In comparison to existing methodologies, suggested model performs
better. Multiple disease prediction system for rice leaf employing better deep leaning techniques will be
included in future scope of this suggested model.

CONFLICT OF INTEREST
The authors declare that they have no conflicts of interest.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023

 DETECTION AND PREDICTION OF RICE LEAF DISEASE USING 57

REFERENCES
1. Kitpo, N. and Inoue, M., Early rice disease detection and position mapping system using drone and IoT architecture,
in 2018 12th South East Asian Technical University Consortium (SEATUC), IEEE, 2018, vol. 1, pp. 1–5.
2. Sharma, M., Kumar, C.J., and Deka, A., Early diagnosis of rice plant disease using machine learning tech-
niques, Arch. Phytopathol. Plant Prot., 2022, vol. 55, no. 3, pp. 259–283.
3. Upadhyay, S.K. and Kumar, A., A novel approach for rice plant diseases classification with deep convolutional
neural network, Int. J. Inf. Technol., 2022, vol. 14, no. 1, pp. 185–199.
4. Nettleton, D.F., Katsantonis, D., Kalaitzidis, A., Sarafijanovic-Djukic, N., Puigdollers, P., and Confalonieri, R.,
Predicting rice blast disease: Machine learning versus process-based models, BMC Bioinf., 2019, vol. 20, no. 1,
pp. 1–16.
5. Sethy, P.K., Barpanda, N.K., Rath, A.K., and Behera, S.K., Deep feature based rice leaf disease identification
using support vector machine, Comput. Electron. Agric., 2020, vol. 175, pp. 105527.
6. Chen, W.L., Lin, Y.B., Ng, F.L., Liu, C.Y., and Lin, Y.W., RiceTalk: Rice blast detection using Internet of
Things and artificial intelligence technologies. IEEE Internet Things J., 2019, vol. 7, no. 2, pp. 1001–1010.
7. Agrawal, M.M. and Agrawal, D.S., Rice plant diseases detection and classification using deep learning models:
a systematic review, J. Crit. Rev., 2020, vol. 7, no. 11, pp. 4376–4390.
8. Liang, W.J., Zhang, H., Zhang, G.F., and Cao, H.X., Rice blast disease recognition using a deep convolutional
neural network, Sci. Rep., 2019, vol. 9, no. 1, pp. 1–10.
9. Bari, B.S., Islam, M.N., Rashid, M., Hasan, M.J., Razman, M.A.M., Musa, R.M., and Majeed, A.P.A.,
A real-time approach of diagnosing rice leaf disease using deep learning-based faster R-CNN framework, PeerJ
Comput. Sci., 2021, vol. 7, e432.
10. Rahman, C.R., Arko, P.S., Ali, M.E., Khan, M.A.I., Apon, S.H., Nowrin, F., and Wasif, A., Identification and
recognition of rice diseases and pests using convolutional neural networks, Biosyst. Eng., 2020, vol. 194, pp. 112–
120.
11. Jiang, F., Lu, Y., Chen, Y., Cai, D., and Li, G., Image recognition of four rice leaf diseases based on deep learn-
ing and support vector machine, Comput. Electron. Agricult., 2020, vol. 179, p. 105824.
12. Shrivastava, V.K., Pradhan, M.K., Minz, S., and Thakur, M.P., Rice plant disease classification using transfer
learning of deep convolution neural network, Int. Arch. Photogramm., Remote Sens. Spat. Inf. Sci., 2019, vol. 42,
no. 3, p. W6.
13. Atole, R.R. and Park, D., A multiclass deep convolutional neural network classifier for detection of common
rice plant anomalies, Int. J. Adv. Comput. Sci. Appl., 2018, vol. 9, no. 1, pp. 67–70.
14. Krishnamoorthy, N., Prasad, L.N., Kumar, C.P., Subedi, B., Abraha, H.B., and Sathishkumar, V.E., Rice leaf
diseases prediction using deep neural networks with transfer learning, Environ. Res., 2021, vol. 198, p. 111275.
15. Larijani, M.R., Asli-Ardeh, E.A., Kozegar, E., and Loni, R., Evaluation of image processing technique in iden-
tifying rice blast disease in field conditions based on KNN algorithm improvement by K-means, Food Sci. Nutr.,
2019, vol. 7, no. 12, pp. 3922–3930.
16. Kumar, R., Baloch, G., Pankaj, A.B.B., and Bhatti, J., Fungal blast disease detection in rice seed using machine
learning, Int. J. Adv. Comput. Sci. Appl., 2021, vol. 12, pp. 248–258.
17. Alwazzan, M.J., Ismael, M.A., and Ahmed, A.N., A hybrid algorithm to enhance colour retinal fundus images
using a Wiener filter and CLAHE, J. Digital Imaging, 2021, vol. 34, no. 3, pp. 750–759.
18. Sahu, S., Singh, A.K., Ghrera, S.P., and Elhoseny, M., An approach for de-noising and contrast enhancement
of retinal fundus image using CLAHE, Opt. Laser Technol., 2019, vol. 110, pp. 87–98.
19. Zhang, S., Zhao, Y., and Bai, P., Object localization improved grabcut for lung parenchyma segmentation, Pro-
cedia Comput. Sci., 2018, vol. 131, pp. 1311–1317.
20. Senthilkumar, C. and Kamarasan, M., A novel citrus disease detection and classification using deep learning
based inception resnet v2 model, Adalya J., 2020, vol. 9, no. 2. http://adalyajournal.com/.
21. Dey, A. and Bandyopadhyay, S.K., Automated glaucoma detection using support vector machine classification
method, Br. J. Adv. Med. Med. Res., 2016, vol. 11, no. 12, pp. 1–12.
22. https://www.kaggle.com/datasets/chandrug/riceleafdisease.
23. Malathi, K. and Nedunchelian, R., Efficient Method To Detect And Classify Diabetic Retinopathy Using Ret-
inal Fundus Images, International Journal of Pure and Applied Mathematics, 2017, vol. 116, no. 21, pp. 89–97,
ISSN: 1311–8080.
24. Malathi, K. and Nedunchelian, R., A recursive support vector machine (RSVM) algorithm to detect and classify
diabetic retinopathy in fundus retina images, Biomedical Research, Volume Special Issue, pp. 1–8, ISSN: 0970-
938X.
25. Malathi, K. and Nedunchelian, R., An Automated Detection of Optic Disc From Digital Retinal Fundus Im-
ages Using Region Based Segmentation Technique, International Journal of Applied Engineering Research,
vol. 10, Issue 1, ISSN: 0973-4562.
26. Malathi, K. and Nedunchelian, R., An Efficient Method to Detect Diabetic Retinopathy Using Gaussian-Bi-
lateral and Haar Filters with Threshold Based Image Segmentation, Research Journal of Applied Sciences, Engi-
neering and Technologу, 2014, 8(11), 1389–1395, ISSN: 2040-7459.
27. Malathi, K. and Nedunchelian, R., Comparison of Various Noises and Filters for Fundus Images using Pre—
processing Techniques, International Journal of Pharma and Bio Sciences, vol. 5, Issue 3, ISSN 0975-6299.

OPTICAL MEMORY AND NEURAL NETWORKS Vol. 32 No. 1 2023