Page 54 - Kaleidoscope Academic Conference Proceedings 2024

2024 ITU Kaleidoscope Academic Conference

(CNN) Deep Net. undergone classification of the diseased and healthy
leaves using a conventional CNN classifier. They have
In this present work, three different models were endured augmentation process for the increase of the
implemented for the classification purpose: a custom data samples. Anushka et. al [6] proposed the idea of
designed fine-tuned 5-layer CNN, a 4-layer CNN, and
MobileNet. These models were used as the backbone again classification with the CNN model and obtained
for the detection model. Among the three models, the an accuracy of 91.41%. A comparative analysis for the
proposed custom 5-layer CNN exhibited the highest classification of potato leaf diseases with the existing
accuracy, achieving nearly 97%. Additionally, a models viz. Googlenet, Resnet and VGG16 is gone
comparative analysis of the accuracies was conducted through and obtained 97% model accuracy [7].An
with different optimizers, including Adam, AdamW, effective idea with amalgamation of classical machine
and SGD.
learning and deep learning is proposed in [8]. Authors
For the detection [1] purpose of potato leaf diseases build the model with K-means clustering segmentation
Faster R-CNN, a state-of-the-art object detection followed by VGG-16 for the final classification of the
algorithm, was implemented. Three categories of potato leaf disease. An accuracy of 97% is achieved. A
potato leaves were considered: Early blight, Late 14-layer CNN is utilized for the purpose of
blight, and healthy leaves. These categories were classification along with augmentation of the data
extracted from the Plant Village dataset, a samples[9]. A 5-class classification on different types
comprehensive database of plant images for research
and development. of potatoes is carried out in the paper[10] with a CNN
model and compared with other models such as
The best results were yielded by the detection model Alexnet, Googlenet, VGG, R-CNN. Segmentation with
when using the 5-layer custom CNN model as the the K-means clustering is used for feature extraction
backbone, generating minimal model loss. and multi-class support vector machine is implemented
Additionally, the outputs generated after detection can for the classification purpose[11]. The accuracy
be easily visualized by the user, providing information obtained with the kind of classical machine learning
such as category and score. The best Intersection over
Union (IoU) was obtained with the 5- layer CNN as the technique is 95.99%. All the existing works are
backbone to the Faster R-CNN, achieving an IoU of undergone majorly with the classification task and
0.76 with a threshold of 0.6. This indicates the some of the works are with the combination of classical
robustness of our detection model in accurately machine learning algorithms. In the present work, a
identifying and localizing potato leaf diseases. novel proposal with a customized CNN model along
with Faster Region based CNN(FRCNN) is executed
The organization of the remaining paper is as follows:
Section 2 exhibits the literature review, Section 3 for the detection of diseased potato leaf and the model
highlights the overall methodology, Section 4 show up generated effective results in terms of the performance
the result and discussion and finally Section 5 parameters.
summarize the entire work with the conclusion.
3. METHODOLOGY

2. RELATED WORKS
The entire workflow of potato leaf disease detection is
depicted in Fig. 1. The overall work has been processed
An exhaustive study of the existing literature is carried using plant Village dataset [12] available at kaggle
out for the present work and are discussed in this repository. It comprises images of healthy and diseased
section with the key points. Sofuoglu C I and Birat D in leaves of various plants out of which potato plant leaves
[2] proposed a deep learning model for classifying the data were taken. There are 3 categories of leaves-Early
plant diseases. The CNN model used, acted as filter to blight, Late blight and healthy. The dataset is computed
the input images for extraction of the key features and using 3 different architectures viz. Mobile net, 4-layer
thereafter classification is done. In [3] authors custom CNN and 5-layer custom fine-tuned CNN. The
implemented a simple CNN classifier for the Accuracy graphs are generated for the three respective
classification of the potato leaf diseases with some architectures. The training is followed by analysis of
basic image pre-processing. A pre-trained model with classification reports for a comparative study amongst
the fine tuning of the hyper-parameters is implemented the different parameters of the respective architectures.
[4]. The model after training acquired 97.4% of In the next step, detection was carried out where in the
classification accuracy for the categories of early blight classified outputs are fed in to the detection
and late blight. Islam and Sikdar[5] in their article architecture, which in our case is F-RCNN, to get

– 10 –