Description: “TOMCROP: A DEEP LEARNING ARCHITECTURE FOR TOMATO LEAF DISEASE DETECTION”

FIELD OF THE INVENTION

The present invention in general relates to a system and a method to detect tomato leaf disease at an early stage and more particularly, to the method for the identification of tomato leaf disease using pre-trained models which includes feature extraction and classification. The experimental results and comparative analysis demonstrate that the system outperforms existing techniques for detecting tomato leaf disease.

BACKGROUND OF THE INVENTION

Tomatoes are an important crop in mineral salt and vitamin C-rich agricultural areas. Plant health is important for crop output growth in terms of quantity and quality. However, to meet these goals, some significant preventive measures need to be adapted so as to detect infections at an early stage.
Tomato crop security is a prime concern with the growth rate of the population. Its high nutrition's value, the good taste makes it popular. Tomatoes are used all over the world but are particularly prevalent in hot and humid climates. Deep learning plays a tremendous role to achieve by classifying the leaves disease beforehand so that it can benefit the small farmers who generally lack the various preventive and control methods to protect their crops from getting spoiled. The leading producers of tomatoes in the world are China, India, Spain, the USA, Brazil, Egypt, Turkey, Iran, Italy. After onion and potato, it is the most cultivated vegetable crop in the world, thus, contributes a lot towards the economy as well. Its growth rate should be improved with the growing population rate [4,11,41]. For economic and generational growth, it should be required in very high quantities. It has been increased due to the continued expansion of greenhouse tomato growing areas. Plant's growth is vulnerable to diseases [27,41]. It has risen as the greenhouse tomato producing area is expanded continuously. These diseases have developed recently. Due to the lack of methods, timely diagnose of the diseases, the farmers are not able to prevent their products from being getting ruined [43]. The challenges and the issues involved in the plant leaf disease detection are:
Detection of multiple infection in the single or multiple leaves,
Computational time complexity involved,
The overlapping of the surrounding stems and their shades in identifying the diseases,
Backgrounds are complex with different real-time environments, and
Intensity of the infection in the leaves.
So, if monitoring of the plant leaf images can be done at the early stages, it can help in maximizing the product quality; thus, a real-time and loss-free and validated technology is required to achieve [36]. In this hi-tech era, mobiles are used very popularly and images are also playing a major role. So, this increases the worldwide smartphone into computer vision and image processing. Ashqar et al. [3] discussed image-based leaves detection using deep learning on total 9000 healthy and infected images [23]. On a worldwide scale, they proposed a concept for smartphone-assisted crop disease diagnostics. Disease identification in traditional methods was a door to door visiting approach. The smartphone and computer vision emergencies are quality and time-saving tools to identify the quality growth of the plant. Thus, plant breeding, horticulture, and boosting fungicide efficiency benefit from a quick calculation of disease severity, disease incidence, and disease quality and quantity. Convolutional Neural Networks (CNN) is applied on image classification popularly known as ILSVRC ImageNet large scale visual recognition competition [16,18,31] consisting of 8 layers. Alex Net outstandingly performs in comparison to traditional computer vision algorithms [34]. Thus, Deep Neural Network (DNN) models were introduced with improved accuracy. Object detection was done with two methods: one-stage was named as faster Regional CNN (R-CNN), region-based as two-stage CNN methods relations were found. Ren et al. [32] worked with RPN fully connected CNN scoring well due to object position [31,32]. Creating high-quality bids Fast - CNN will employ a Recurrent Neural Network (RNN) that was trained during complete process. Region Proposal Network (RPN) and Fast R-CNN were also merged into a unified network having shared convolution features. Mohanty et al. [25], worked on an incomplete data set. The method was also applicable to a large number of classes and solved the issues that older methods had. In conventional benchmark problems, it was attained by a significant margin and also proven that without feature engineering the model can correctly classified images. As classification is much faster than training, it can be implemented on smartphones. Transfer learning based deep CNN is designed and tested on pre-trained models Alex net, Google Net, and Res Net [46,47]. These results were compared with the Stochastic Gradient Descent (SGD) method and Adam Optimization observed that Resnet SGD performs the best accuracy. They claimed that a small batch size and few iterations can increase the target model's accuracy for specific jobs. Layer-wise fine-tuning obtained good accuracy for tomato leaf disease. Due to higher computational power in images, deep learning, machine learning has extensively played in agriculture.
Pre-trained Deep learning models
Many authors (Szegedy, 2016; Ioffe 2015; Chollet, F. 2017; Harte, E. 2020; Zhang K, Wu Q, Liu A, Meng X 2018) provide different model versions: Inception (V1 to V4) and ResNet Inception V3 network is designed by Keras, which is trained in ImageNet. Due to the division of large volume integrals, small convolutions make it optimal over inception V1 and V2. The inception architecture is designed with three varied sizes of convolution and a maximum of one pooling layer. The non -linear structure is formed with a network output layer aggregated with convolution operation. This removes the overfitting problem with the use of varied scales. Howard et al. [20] devised a method for reducing the size of deep neural networks that was both simple and effective. The MobileNet architecture is framed as a separable depth-wise convolution type architecture. When determining output, accuracy and latency are taken into account. It may be used for a variety of tasks, including fine-grain classification, facial characteristics, and large-scale geo localisation. TensorFlow [1] was used to train MobileNet models using RMSprop [5,40] and asynchronous gradient descent, comparable to Inception V3 [40,42]. They utilised fewer regularisation and data augmentation strategies when training small models than when training large models because these models were less prone to overfitting. They did not employ side heads or label smoothing when training MobileNets, and they also minimised the number of distortions by restricting the size of small crops used in big Inception training [40]. Furthermore, because the depth-wise filters had so few parameters, it was discovered that very little or no weight decay (l2 regularisation) was required. All models, regardless of size, were trained using the same training parameters for the ImageNet benchmarks in the Section 3. Initially, it simply considered size, therefore inception models were included to speed up the computation. He, K. at al. (2015) hypothesised a residuals link for object detection and image recognition. The training speed is substantially increased because of the residual connections. It popularly allows CNN to be used. DCNN residual connections are necessary for training the data. Highly deep convolutional networks have recently proven essential in the advancement of image recognition capability. The performance of the Inception architecture has been demonstrated at minimal computational costs. Inception-v4 is a mix of inception modules and residual connections that is an enhanced version of Inception-v3. Batch normalisation is performed on top of standard convolutional layers in inception-v4. A popular exception inception architecture designed as an enhanced version of Inception V3 [8]. It improved the performance of the Image dataset which has been analysed with Inception V3. Its performance was verified with 17,000 on 350 million images. The inception V3 and Xception have the same architecture. The Xception is better due to the perfect usage of parameters and 36 Layers with 14 modules. Except for the first and last modules, the rest are bounded with linear residual connections. Inception vs Xception weight decay, 4e-5 and 1e-5, drop out is same in both architecture 0.5 before logistic regression, the auxiliary loss is optional, but not included in Xception and at last is parameter count of 23,626,728 in Inception V3 while 22,855,952 in Xception. Huang et al. [21] discussed an efficient architecture, where input and output layers relate to short connections. The main pros of this type of network are the good flow of gradients and information in all-around networks [21,42]. Further, performance is increased consistently, without any overfitting and degradation. So, these can be performed well with fewer parameters and less computation. VGG16 is a 16-layer CNN model developed by researchers from the University of Oxford. KIt achieves an accuracy of over 92.7% on ImageNet, a huge dataset with over 14 million images covering 1000 identities. It took the developers weeks to train VGG16, even with NVIDIA Titan Black GPUs. The network weights data is beyond 533MB. Due to these limitations, it is difficult to train on conventional hardware. Pre-trained versions can however be used for learning and experimentation. VGG19 is an upgraded version, with 19 layers [2]. Deep learning models aim to concede efficient models with smaller framework. Since 2012, success in the ImageNet database has become little harder, but some of them are effective in context of computing load. Among the state-of-art model, Efficient Net model reaches to good accuracy with 66M parameter in ImageNet classification problem that may be considered group of CNN models. Effective outcomes with uniform scaling of width, depth, and resolution rate also. A compound scaling method is employed to find the integrity in different dimension with fixed constraint. These co-efficient constants are deployed to find the integrity in different dimensions with fixed constraint. The Baseline B0 model is enhanced up to B7 uses the same scaling methods. Basic required resources are considered double with the optimized parameters. Efficient Net consists of 8 models during B0 and B7, and as higher version model appears, number of parameters are not as much increased as accuracy increases rapidly. It works on swish activation function instead of Relu activation function [49]. It was proposed in Mobile Net V2, but due to higher number of FLOPS (floating point operation per second) Efficient Net leads the results [50]. In MB conv, layers are in the form block expansion more than MobileNet V2, layers are connected directly to block compressed and expanded channels with expansion layers. This advancement has in-depth connections to reduction by squared factor with the traditional layers with kernel size denoting the width and height of 2D convolutional window [50].
Leaf disease detection using deep learning models
Brahimi et al. [7] discussed many deep learning frameworks available like Python Library and R-FCN architect was proposed by Dai et al. [9] for accurate and efficient leaf disease diagnosis. In object detection, position-sensitive score maps were used. This would remove the dilemma between translation variance and invariance which makes it 2.5-20 X faster than faster R-CNN. Fuentes et al. [14] had designed a tomato disease and pest recognition detection technique [10,19]. It was based on the idea of localization and diagnosis of various pests and diseases which is good in handling their own image data set. It explored numerous ways how a plant is impacted by different laboratories and also consider images with different capturing devices at different resolutions. It also takes into account various lighting situations, background differences, and item sizes. The few parameters were required to train the number of samples, trainable models and hyperparameters were used in deep neural networks. Neural Networks were attracted towards overfitting for large data sets [33] to enhance performance, data augmentation increases the input samples or batch normalization [5,35,45], randomly dropout activation functions [37] or weight regularization to decrease architecture overfitting [45]. These methodologies are effectively applied to large networks, incomplete data imbalanced class problems were not solved. While maintaining a small computational cost, deep learning methods can be performed on complex tasks in well manner. Fuentes et al. [15] discussed the leaf infection status denoting different variations of causes of his previous work [14]. He achieved good results with transfer learning and data augmentation of pre-trained models in huge data sets [13]. A refinement filter bank framework was used to solve the class imbalance problem and false-positive problem for tomato plant diseases. CNN based diagnosis unit filter bank is used for verification. For each class, CNN classified filtered misclassification samples independently. The resultant disease category was detected or not based on performance compared to previous work, but it worked with limited data. For disease leaves classification, Balakrishna et al. [6] offered Gray Level Co-occurrence Matrix (GLCM), K-nearest neighbours (KNN), and Hue Saturation Value Format (HSVF), Morphological Operation (MO), Sobel Edge Detection (SED), Gabor Filter (GF), Probability Neural Network (PNN). KNN and PNN classification outperforms KNN classification, according to their findings. It occurred as a result of the PNN's radial basis capabilities. The structural picture pixels describing a neighbourhood are converted using the morphological technique. To extract features from the photos, erosion and dilation were applied to pixels of all neighbourhood images, which also aids in data reduction. The Gabor functionality defines how the joint uncertainty in space and frequency is decreased, resulting in the best possible result. The complexity of CNN architecture is being overcome with the enhanced Crossover Optimization (ICRMBO) algorithm [26,27]. CNN eliminates manual work via encoding methods. Inception V3 and VGG16 were occupied, and ICRMBO was used to optimise them. These methodologies can be used in different applications. Gonzalez-Huitron [16] used transfer learning to train MobileNet V2, NasNet Mobile, Xception, and MobileNetV3 with average pooling layers were followed by a SoftMax layer with 10 epochs, 10 classes, Adam Optimizer with batch size 24, at 0.001 learning rate, 0.9 and ß1 as ß2 at 0.999, ? 1e-7 [16]. These hyperparameters were trained with weight metrics for state-of-the-art and trained models. Ferentinos et al. [12] proposed a plant disease detection model using the CNN architecture on the database having 25 different plants. The database was trained on the models AlexNet, AlexNetOWTBn, GoogleNet, and VGG outperformed all the other model's classification rates 99.48%. The leaf diseases categorization of plants with CNN architecture was introduced [22]. They operated on three distinct plant datasets: tomatoes, peppers, and potatoes, after training and testing has an average accuracy rate of 98.29% and 98.02%, respectively. The prediction model for tomato leaf disease with transfer learning and Deep Convolutional Neural Network to achieve better classification accuracy was introduced in [36]. Thenga raj et al. [41] also used transfer learning and compared the performances of the different optimizers Adaptive Moment Estimation (Adam), RMSprop and Stochastic Gradient Descent (SGD) optimizers. They analysed that Adam Optimizer proves to show better accuracy as compared to the other two specified in [44,49]. They worked on the modified Xception model based on Adam optimizer and measured the consistency, reliability of the model using the performance metrics like recall, precision and F1 score. Rangarajan et. al [28] performed the disease classification in the crops with the six pre-trained models AlexNet, GoogLeNet, DenseNet201, VGG19, Visual Geometry Group 16 (VGG16), and ResNet101 automatically. All the models GoogLeNet proved to give the best validation accuracy and VGG16 gave the best test accuracy. A total of ten diseases were included in the dataset. Tobacco Mosaic Virus (TMV), Epilachna beetle, Two-spotted spider mite, Cercospora leaf spot, Brown spot, Citrus Hindu mite, Citrus canker, Yellow Vein Mosaic Virus, Yellow Vein Mosaic Virus, Yellow Vein Mosaic Virus, Yellow Vein Mosaic Virus, Yellow Vein Mosaic Virus, Yellow Vein Mosaic Virus, Yellow Vein Mosa for four distinct crops, Ladyfinger, Eggplant, Hyacinth Beans, Lime. The EfficientNet deep learning model is evaluated against the other state of art models for the Plant Village Dataset [4]. They performed the experimentation on the augmented and original dataset. Their outcome tells the average accuracy and the performance of the B4 and B5 models is much over the CNN architectures. A deep ensemble transfer learning neural network has analysed the outcome on the Plant Village dataset having 38 classes from 14 crops [44]. DENN is performed much better in comparison to pre-trained models MobileNetV3, InceptionV3, DenseNet 121 & 201, ResNet 50 & 101 and NasNet. With cascading the AlexNet with CNN, GoogLeNet and inception model a novel framework was introduced [30]. The model has been compared with other models like AlexNet, VGGNet-16, GoogLeNet and ResNet-20. The experimentation was carried out on the Apple leaf's datasets and it was stated that DLDPF performed well to predict diseases automatically. Guo et al [17] proposed a plant disease identification model for smart farming in which first of all the recognition and localization of the leaves is done using the Region Proposal Network and then the images are segmented with the Chan-Vese (CV) algorithm. The segmented leaves are then given as the input to the transferring model. The model shows an accuracy rate of 83.57%. Harte et al. [18] empirically evaluated the outcomes of a pre-trained ResNet34 model for detecting disease. This model acted as a web application for the recognition of 7 plant diseases. The model achieved the validation of an F1 score of 96.5% and a classification rate of 97.2%. In past years, Sensor based techniques, infra variations with innovative technology, Drone Technology also led to the identification of disease detection and market sight changes. Drone works with an infrared camera, sprayed pesticides. It shows the plant growth regularly to farmers. The different methodologies it has a new impact. In Drone methodology, image capturing is followed by classification and deep learning algorithms. Appropriately pesticides are sprayed in the required amount [38,43]. The related work is summarized in Table 1:
Table 1: Summary of Leaf Disease Detection Method with Classification Accuracy
Authors Year Method Dataset No. of classes Classification accuracy
Thangaraj et al. [41] 2021 Modified Xception with Adam Optimizer 16,578 10 99.55%
Guo et al. [17] 2020 RPN 1000 3 83.57%
Harte E. [18] 2020 Resnet34 - 97.2%
Ferentinos et al. [12] 2018 ANN 87,848 58 99.48%
Aravind Krishnaswamy
Rangarajan et al. [28,29] 2018 VGG16, AlexNet 13,262 7 97.29%,97.49%
Keke Zhang et al. [46] 2018 GoogLeNet, ResNet, AlexNet 5,550 9 95.66%,97.28%, 95.83%
Durmuú et al [11] 2017 SqeezeNet, AlexNet 19,742 10 94.3%, 95.65%
Mohammed Brahmini et al. [7] 2017 GoogLeNet
AlexNet 14,828 9 99.18%,
98.6%
OBJECTIVES OF THE INVENTION
The prime objective of the present invention is to provide a system and a method to detect tomato leaf disease at an early stage.
Another objective of the present invention is to provide a methodology for the identification of tomato leaf disease using pre-trained models which includes feature extraction and classification. The methodology involves the use of the modified pre-trained model (MobileNet) as the convolutional base for the feature extraction. The model is then hyper-parameter tuned using the various parameters to obtain the best results for classification.
Another objective of the present invention is to provide a method to detect the details such as an infection stage and identifying symptoms that are not visible to most human observers.
Another objective of the present invention is to provide a method to investigate the disease resistance qualities of new crop producers in the lab.
Another objective of the present invention is to provide a method for appropriate diagnosis of infection at the early stages which leads to more economical use of pesticides and leads to less cost of production.
These and other objectives of the present invention will be apparent from the drawings and descriptions herein. Every objective of the invention is attained by at least one embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings described herein are for illustrative purposes for selected embodiments only and not all possible implementations. The amount of detail offered is not intended to limit the scope of the present disclosure. Further objectives and advantages of this invention will be more apparent from the ensuing description when read in conjunction with the accompanying drawing and wherein:
Figure 1 represents a deep learning architecture with a stack of fully-connected layers followed by a Softmax layer as classifiers.

Figure 2 represents the graphs of analysis of the validation and training loss vs epoch for different pre-trained models used as feature extractor in the method.
Figure 3 represents the graphs of analysis of the validation and training accuracy vs epoch for different pre-trained models used as feature extractor for the method.
Figure 4 represents the confusion matrix using the ‘Adam’ optimizer for the method with all the pre-trained models as feature extractors.
Figure 5 represents the training and validation accuracies graph for MobileNet for various optimizers.

Figure 6 represents the comprehensive framework of the method.

It will be recognized by the person of ordinary skill in the art, given the benefit of the present disclosure, that the examples shown in the figures are not necessarily drawn to scale. Certain features or components may have been enlarged, reduced, or distorted to facilitate a better understanding of the illustrative aspects and examples disclosed herein. In addition, the use of shading, patterns, dashes, and the like in the figures is not intended to imply or mean any particular material or orientation unless otherwise clear from the context.

DETAILED DESCRIPTION OF THE INVENTION
The embodiments herein and the various features and advantageous details thereof are explained more comprehensively with reference to the non-limiting embodiments that are detailed in the following description. Descriptions of well-known components and processing techniques are omitted so as to not unnecessarily obscure the embodiments herein.
The following description includes the preferred best mode of one embodiment of the present invention. It will be clear from this description of the invention that the invention is not limited to these illustrated embodiments but that the invention also includes a variety of modifications and embodiments thereto. Therefore, the present description should be seen as illustrative and not limiting.
Unless otherwise specified, all terms used in disclosing the invention, including technical and scientific terms, have the meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, term definitions may be included to better appreciate the teaching of the present invention.
As used in the description herein, the meaning of “a,” “an,” and “the” includes plural reference unless the context clearly dictates otherwise. Also, as used in the description herein, the meaning of “in” includes “in” and “on” unless the context clearly dictates otherwise.
One embodiment of the present invention provides a system that includes an input device, means for data pre-processing, pre-trained model module, modified layers module, and a unit of analysis.
Another embodiment of the present invention is a method to detect disease in tomato leaves which includes the steps as collecting a dataset of the tomatoes and pre-processing the dataset, using pre-trained models as a convolutional base for feature extraction of the tomato dataset, feeding a stack of fully connected layers as a classifier by the features extracted from the pre-trained models, hyperparameter tuning of the models, analyzing and comparing the pre-trained models, and obtaining the results based on the compared data.
The method of the present invention includes the identification of Tomato leaf disease using pre-trained models by feature extraction and classification. In the method, there are two different models. In the first model, the pre-trained models (Inception3, MobileNet, Resnet50, Xception, Densenet121, VGG16 and EfficientNet B0) are used as the convolutional base for feature extraction and then a stack of fully connected layers followed by a SoftMax layer as classifiers that are fed by the features extracted from the pre-trained models. In the second model, TomCrop: A deep learning architecture is used which involves fine-tuning of the pre-trained model MobileNet by unfreezing some of the layers of the convolutional base and retraining the network with a lower learning rate and an optimizer. Through empirical analysis it is observed that the MobileNet model performs better than the remaining models, thus the hyperparameter tuning of the model with MobileNet as feature extractor is also done for different optimizers Adam, SGD, Adagrad, Adadelta and RMSprop and the results have been analysed. Also, the experimentation of the MobileNet model for the various batch sizes 32, 64 and 128 has also been done. The results obtained from the deep learning architectures are then compared in terms of precision, recall, F1 score. Comparative analysis and the experimental results verify the efficiency of the method with existing systems for tomato leaf disease detection.
METHODOLOGIES
Deep learning architectures are capable of solving a wide range of applications that seems impossible to be solved earlier, although, the problems associated with deep learning are of large training data, need much time to train the data and occupies a large cluster of computers or servers.
DATASET USED
The “Tomato” dataset from Kaggle website has been used for the experimental evaluation. The dataset consists of 22930 images which is split into two categories: training data and test data. The dataset comprises 10 different diseases and the sample images of the leaves having these diseases are arranged in 10 different folders. The dataset could be useful for the tomato leaf disease classification which could prove to be beneficial for the farmers as it can prevent crop failure and hence the economic loss. To perform training, the model is trained with a training dataset and its performance is validated using the test images. Table 2 depicts some images of the different categories from the tested dataset.
Table 2: Some of the sample images from the Tomato Dataset

Healthy Bacterial Spot Early blight Late Blight Leaf Mold Septoria
Leaf Spot Spider Mites Target Spot Tomato Mosaic Virus Tomato Yellow Leaf Curl Virus
s

TESTING AND TRAINING DATASET
The complete dataset is grouped in training and tested data with an 80:20 ratio. The training folder has 18345 images, and the test folder has 4585 images. Image sizes are equally distributed in the form of size with 256 × 256 pixels. Each image in the testing dataset as well as the training dataset is reformed with the size of 224 × 224 pixels to fit the model. The fine-tuning of the model is done by setting the learning rate to 0.001, epochs to 15 and the batch size as 32 to train the network model. It is trained with the 'Adam' optimizer.

DATA PRE-PROCESSING
Data normalization and augmentation are crucial steps that are needed for the pre-processing of the data before training images with deep learning models. Pixel values in the images range from 0 to 255 and these values need to be scaled before feeding them into the deep learning neural network while training or testing the model. Keras uses an important class "ImageDataGenerator" for scaling the images. ImageDataGenerator supports three types of pixel scaling techniques: the scaling of the pixel is done in the range of [0-1] is Pixel Normalization (PN), scaling these pixel values with mean zero is Pixel Centring (PC) and the scaling of these pixel values with unit variance and mean zero is Pixel Standardization (PS). Normalization not only moderates the internal covariate shift but makes the optimization faster and also supports the network in regularization. Data augmentation is also one of the important steps for data pre-processing. Its efforts to increase the dataset by introducing diversity and variety in the dataset when there is a limited dataset by introducing various types of transformations. Data augmentation helps to resolve the overfitting problem. This happened when the training accuracy is much greater than the testing accuracy i.e., the neural network model learns the dataset while doing training and hence does not perform up to mark when during testing some unknown input is given to the model to analyses the model performance. Another way to prevent the problem of overfitting is a dropout. Dropout is a regularization technique in which certain neurons are removed at each training step.
TESTING AND TRAINING MODEL
The model learns automatically by allowing the network to be trained using the large image dataset on the powerful and robust GPUs with DCNN. The experiments are analysed on the 11th generation Intel Core i5 2.40 GHz with 8GB RAM. The implementation of the model is done using python programming with GPU on the dataset from the Kaggle website. Once the data pre-processing is done the features are extracted using convolution layers and pooling layers from the given images. Activation functions are one of the important building blocks of the neural network which can add non-linearity into a network. The choice of the activation function controls the performance of the model. In this paper, “SoftMax” and Rectified Linear Unit (“ReLu”) activation functions have been used as a generalized form of the activation function “Sigmoid” for multi-class models.

Introducing the first model which involves the use of a stack of fully-connected layers as classifiers that are fed by the features extracted from the pre-trained models (Inception3, MobileNet, Resnet50, Xception, Densenet121, VGG16, EfficientNetB0). Here, the pre-trained models are used as the convolutional base for feature extraction and then a new densely connected classifier is added on top of the pre-trained model to train it on the extracted features. The model having the weights which were obtained by training the ImageNet dataset is loaded by Kera’s. The model is loaded excluding the densely connected classifiers. The layers of the pre-trained models are frozen i.e., the weights are made non-trainable, and the gradient descent will not update them and then a simple fully-connected classifier is added using the Sequential model. Firstly, a data augmentation model is added, then a Rescaling layer is added to standardize the values to [0, 1] range, followed by the pretrained models as convolutional base, and then finally a new fully-connected classifier is added. The Dense layer have 10 nodes as their output and since the problem being dealt here is a multiclass classification so, the SoftMax activation function is used.

Figure 1 shows the complete framework of the present method. It represents a deep learning architecture with a stack of fully-connected layers followed by a Softmax layer as classifiers. Another, most used activation function that is quite easy to compute and does not have the problem of saturation and also does not cause the Vanishing Gradient Problem in CNN is the "ReLu".

It is used as an activation function in the hidden layer and the choice of the activation function for the output layer depends upon the type of the prediction problem. Thus, to underrate the loss function and to make it correct and optimized predictions as possible changes need to be made in the weights and other parameters like the learning rate of the model. Here optimizers play the major role. There are different types of optimizers like Adagrad, RMSprop, Adam. In the models "Adam" optimizer has been used which stands for "adaptive moment estimation. It is one of the most popular neural network optimizers. The metrics values are captured after every epoch on the training dataset. The seven pre-trained models Resnet50, InceptionV3, Xception, Densenet121, Mobilenet, VGG16, EfficientNet are empirically implemented using the GPU on the tomato dataset. To address the problem of multi-class classification tasks "accuracy" metric has been used to report the accuracy. The training vs validation accuracy and training vs validation loss plots has been shown in figure 2 and figure 3. Figure 2 represents analysis of the validation and training loss vs epoch for different pre-trained models used as feature extractor in the method and figure 3 represents analysis of the validation and training accuracy vs epoch for different pre-trained models used as feature extractor for the method.

Table 3 shows the validation and training loss and accuracy vs 15 epochs for the various pre-trained models from where the model gives the highest training accuracy and the highest validation accuracy for the MobileNet as pre-trained model for feature extraction. Performance indicators like precision, accuracy, recall, and F1 score can be used to assess the model's strength. The ratio of number of accurately predicted photos to the total number of predictions can be used to calculate the accuracy.

Table 3: The training and validation loss and accuracy vs 15 epochs for the various pre-trained models as feature extractor
Model Epochs Train loss Train accuracy Validation Loss Validation accuracy
InceptionV3 15 1.2565 0.9216 2.7146 0.86 96
Resnet50 15 1.2729 0.6537 1.4159 0.6349
MobileNet 15 0.3711 0.9741 2.1767 0.9437
Densenet121 15 0.5098 0.9662 1.2991 0.9398
Xception 15 0.8396 0.9438 2.1049 0.8955
VGG16 15 0.0923 0.9687 0.4259 0.8975
EfficientNet B0 15 0.2340
0.9220 0.2336 0.9336

Classification accuracy rate= (Corrected Predictions)/(Correctected predictions+Non Correctected predictions) (1)

Precision metrics, which can be defined as the proportion of properly corrected results (TP) with the corrected and non-corrected, are another type of performance metric.
Precision= TP/(TP+FP) (2)

The proportion of true positive results (TP) to the total number of samples (TP + FN) is one of the Recall metrics. Recall= TP/(TP+FN) (3)

TP stands for True Positive, FP for False Positive, TN for True Negative, and FN for False Negative. The model performance is calculated using the F1 score, which is calculated by taking the weighted harmonic mean of precision and recall.

F1 score=2× (Recall ×Precision)/(Recall+Precision) (4)
Table 4: The classification report in terms of F1-score, Precision and Recall for the various pre-trained models as feature extractor

DenseNet121 MobileNet VGG16
Classification
Report Precision Rate Recall Rate F1-Score Precision Rate Recall Rate F1-Score Precision Rate Recall Rate F1-Score
Bacterial spot 0.93 0.88 0.90 0.99 0.78 0.87 0.97 0.86 0.91
Early Blight 0.64 0.77 0.70 0.76 0.68 0.72 0.64 0.77 0.70
Late blight 0.78 0.83 0.80 0.81 0.87 0.84 0.85 0.72 0.78
Leaf mold 0.75 0.83 0.79 0.72 0.89 0.80 0.85 0.68 0.75
Septoria leaf spot 0.87 0.75 0.81 0.92 0.74 0.82 0.75 0.89 0.81
Spider mites Two-spotted spider mite 0.75 0.89 0.82 0.71 0.93 0.81 0.66 0.95 0.78
Target Spot 0.77 0.78 0.78 0.64 0.87 0.74 0.66 0.85 0.74
Tomato mosaic virus 0.99 0.86 0.92 0.93 0.98 0.96 0.98 0.77 0.86
Tomato Yellow Leaf Curl Virus 0.92 0.98 0.95 0.97 0.92 0.95 0.86 1.00 0.93
Healthy 0.99 0.69 0.81 0.99 0.52 0.68 1.00 0.45 0.62

InceptionV3 Resnet50 Xception
Classification
Report Precision Rate Recall Rate F1-Score Precision Rate Recall Rate F1-Score Precision Rate Recall Rate F1-Score
Bacterial spot 0.89 0.62 0.73 0.55 0.86 0.67 0.93 0.58 0.71
Early Blight 0.48 0.75 0.59 0.52 0.56 0.54 0.60 0.55 0.57
Late blight 0.43 0.91 0.58 0.65 0.48 0.55 0.65 0.88 0.75
Leaf mold 0.73 0.51 0.60 0.73 0.37 0.49 0.81 0.71 0.76
Septoria leaf spot 0.83 0.26 0.39 0.43 0.53 0.47 0.85 0.55 0.66
Spider mites Two-spotted spider mite 0.62 0.89 0.73 0.94 0.28 0.43 0.55 0.92 0.69
Target Spot 0.67 0.60 0.63 0.71 0.46 0.56 0.59 0.79 0.67
Tomato mosaic virus 0.97 0.78 0.86 0.44 0.97 0.60 0.83 0.89 0.86
Tomato Yellow Leaf Curl Virus 0.89 0.88 0.89 0.88 0.79 0.83 0.94 0.89 0.91
Healthy 0.96 0.45 0.61 0.97 0.72 0.82 0.97 0.54 0.69

To visualise the model's performance confusion matrix is used and is often expressed in tabular form. A confusion matrix is divided into four sections. True positive (TP), true negative (TN), false positive (FP), and false-negative (FN) are the four types of true positives and false negatives (FN). The anticipated output class is shown in columns, whereas the goal class is represented in rows. The confusion matrix using the ‘Adam’ optimizer for all the six models are shown in figure 2. Figure 4 shows the confusion matrix using the ‘Adam’ optimizer for the model with the all the pre-trained models as feature extractors.

To regulate the behaviour of an ML algorithm, hyperparameters are knobs or settings that can be tweaked before performing a training job. They can have a significant impact on model training in terms of training time, infrastructure resource requirements (and thus cost), model convergence, and model correctness. A smart selection of hyperparameters can either help a model attain the intended metric value or lead to an endless cycle of continuous training and optimization. The number of layers, epochs, activation functions, learning rate, and other parameters are among them. There are different types of hyperparameters: Model hyperparameters, Optimizers hyperparameters and data hyperparameters. Generally, two basic approaches are used for the selection of hyperparameters i.e., Grid Search and Random search. The hyperparameter tuning of the model of the present invention with MobileNet as convolutional base is also done for various optimizers Adam, SGD, Adagrad, Adadelta and RMSprop and the results have been analysed. Table 5 given below shows the comparative analysis of the model of the present invention (with MobileNet as feature extractor) using the different optimizers in terms of training and validation accuracy and training and validation loss. Also, the experimentation of the MobileNet model has been for the various batch sizes 32, 64 and 128 has also been done.
Table 5: The comparative analysis of the model of the present invention using the different optimizers
Optimizers Training loss Training Accuracy Validation Loss Validation Accuracy
Adam 0.3711 0.9741 2.1767 0.9437
AdaGrad 0.0576 0.9819 0.1493 0.9533
Rmsprop 0.3577 0.9796 0.4703 0.9490
AdaDelta 0.6356 0.7852 0.6477 0.7788
SGD 0.1651 0.9832 0.6760 0.9577

Figure 5 shows the training and validation accuracies graph for MobileNet for various optimizers. Batch size may be defined as how many samples will be propagated through the network at a time and it is one of the most important hyperparameters which helps to make sure the models hit peak performance. In general, the higher the batch size, the faster the model will complete each training cycle. This is due to the fact that, depending on the computational capacity, the system of the present invention is able to process much more than one sample at a time. The trade-off is that, even if the machine can handle very big batches, the model's quality may decline due to increase in the batch size, causing the model to be unable to generalize successfully on data it hasn't seen before. Through the experimentation it is analyzed that increasing the batch size reduces the validation loss and the training accuracy is also degrading.

Table 6: The comparative analysis of the model of the present invention using the different batch size
Batch Size Training loss Training Accuracy Validation Loss Validation Accuracy
32 0.3711 0.9841 2.1767 0.9437
64 0.4297 0.9720 1.0766 0.9562
128 0.1750 0.9689 0.5566 0.9453

Fine-tuning of the pre-trained model MobileNet is accomplished in the second suggested model by unfreezing the top layers of the convolutional base and retraining the network with a low learning rate and SGD optimizer. The suggested model's comprehensive framework is shown in figure 6. The model's accuracy can be increased by fine-tuning. Unfreezing the few top layers of the convolutional base and retraining the network with a modest learning rate are the main components of fine-tuning.
The frozen pre-trained layers will, as is customary, convolve visual attributes. The pre-trained layers that aren't frozen (i.e., 'trainable') will be trained on the dataset and updated depending on the Fully-Connected layer's predictions. The number of layers removed while fine-tuning a model varies depending on the circumstance, however it has been shown via testing that removing the final 5 layers is sufficient for this purpose. As a result, the vast bulk of the original MobileNet design, which consists of 88 levels in total, is preserved with this arrangement. Finally, a layer named output is added, which is just a Dense layer with ten output nodes for each of the 10 classes, as well as the 'softmax' activation function.
Following the formation of the model, the number of layers to be trained on the dataset is determined. Despite the objective of the present invention is to preserve most of what the original MobileNet learned from ImageNet by freezing the weights in many layers, particularly the early ones, there is still need to train certain layers since the model has to learn the characteristics about this new data set. After some testing, it was discovered that training only the final 23 layers and freezing the rest of the layers can produce acceptable results.
Figure 6 shows the complete framework of the deep learning architecture: TomCrop

Table 7 mentions the number of trainable and non-trainable parameters for the original MobileNet model and the MobileNet model (TomCrop) of the present invention. The MobileNet model of the present invention has 1,873,930 trainable parameters and 1,365,184 non-trainable parameters. The model is run for 15 epochs. The model is compiled using the ‘Adam’ optimizer, loss function as ‘categorical_crossentropy’ and a learning rate of 0.0001. The model gives a training accuracy of 99.97% and the validation accuracy of 97.86% which is quite a significant progress from the first model where the model involved the use of a stack of fully-connected layers as classifiers and all the layers were freezed.
Table 7: Comparative analysis of the models.

Model Total Parameters Trainable Parameters Non-trainable Parameters Epochs Train loss Train accuracy Validation Loss Validation accuracy
Original MobileNet 4,253,864 4,231,976 21,888 15 0.3711 0.9741 2.1767 0.9437
MobileNet (TomCrop) 3,239,114 1,873,930 1,365,184 15 0.0021 0.997 0.0730 0.9786

Conclusions
Plant disease heavily affects the quality and the production which proves to be an unbearable financial loss to the small farmers. However, if it is possible to early detect diseases by the use of some reliable methods and technology it can prove to be a great boon for the farmers as the necessary reforms or measures can be taken to protect the crop from getting ruined. The recent trends show that the advancement in the field of convolutional neural networks and deep learning architectures have given an immense thrust in the problems of image classification, segmentation and object detection tasks. In the present invention, an empirical evaluation of the state of art pre-trained deep neural network architecture has been done to look into the suitability of these models for the classification of the tomato leaf disease problem. In the first model wherein the pre-trained models (Inception3, MobileNet, Resnet50, Xception, Densenet121, VGG16 and EfficientNetB0) are used as the convolutional base for feature extraction and then a stack of fully connected layers followed by a SoftMax layer as classifiers it was concluded that with MobileNet the model gave the best performance of training accuracy of 97.41% and a validation accuracy of 94.37%. Further, the hyperparameter tuning of the model has been done to see the effects of the different optimizers Adam, SGD, Adagrad, Adadelta and RMSprop and the different batch sizes 32, 64 and 128 using the model with MobileNet as Convolutional base. Further, in the second model, TomCrop: A deep learning architecture is used which involves fine-tuning of the pre-trained model MobileNet by unfreezing some of the layers of the convolutional base and retraining the network with a lower learning rate and Adam optimizer. TomCrop achieves a training accuracy of 99.7% and the validation accuracy of 98.7%. The experimental results and comparative analysis demonstrate that the system of the present invention soutperforms existing techniques for detecting tomato leaf disease.
Although the invention has been described with reference to specific embodiments, this description is not meant to be construed in a limiting sense. Various modifications of the disclosed embodiments, as well as alternate embodiments of the invention, will become apparent to persons skilled in the art upon reference to the description of the invention. It is therefore contemplated that such modifications can be made without departing from the scope of the present invention as defined.
REFERENCES
[1] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., & Zheng, X. (2016). Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467.
[2] Alippi, C., Disabato, S., & Roveri, M. (2018, April). Moving convolutional neural networks to embedded systems: the alexnet and VGG-16 case. In 2018 17th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (pp. 212-223). IEEE.
[3] Ashqar, B. A., & Abu-Naser, S. S. (2018). Image-based tomato leaves diseases detection using deep learning.
[4] Atila, Ü., Uçar, M., Akyol, K., & Uçar, E. (2021). Plant leaf disease classification using EfficientNet deep learning model. Ecological Informatics, 61, 101182.
[5] Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer normalization. arXiv preprint arXiv:1607.06450.
[6] Balakrishna, K., & Rao, M. (2019). Tomato plant leaves disease classification using KNN and PNN. International Journal of Computer Vision and Image Processing (IJCVIP), 9(1), 51-63.
[7] Brahimi, M., Arsenovic, M., Laraba, S., Sladojevic, S., Boukhalfa, K., & Moussaoui, A. (2018). Deep learning for plant diseases: detection and saliency map visualisation. In Human and machine learning (pp. 93-117). Springer, Cham.
[8] Chollet, F. (2017). Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1251-1258).
[9] Dai, J., Li, Y., He, K., & Sun, J. (2016). R-fcn: Object detection via region-based fully convolutional networks. arXiv preprint arXiv:1605.06409.
[10] Donatelli, M., Magarey, R. D., Bregaglio, S., Willocquet, L., Whish, J. P., & Savary, S. (2017). Modelling the impacts of pests and diseases on agricultural systems. Agricultural systems, 155, 213-224.
[11] Durmus H, Gunes EO, Kirci M, (2017) Disease detection on the leaves of the tomato plants by using deep learning. In: 6th IEEE International Conference on Agro-Geoinformatics (2017) pp. 1–5.
[12] Ferentinos, K. P. (2018). Deep learning models for plant disease detection and diagnosis. Computers and Electronics in Agriculture, 145, 311-318.
[13] Fuentes, A. F., Yoon, S., Lee, J., & Park, D. S. (2018). High-performance deep neural network-based tomato plant diseases and pests diagnosis system with refinement filter bank. Frontiers in plant science, 9, 1162
[14] Fuentes, A., Carlos, N., Ruiz, Y., Callard, D., Sánchez, Y., Ochagavía, M. E., ... & Pooggin, M. M. (2016). Field trial and molecular characterization of RNAi-transgenic tomato plants that exhibit resistance to tomato yellow leaf curl geminivirus. Molecular Plant-Microbe Interactions, 29(3), 197-209.
[15] Fuentes, A., Yoon, S., Kim, S. C., & Park, D. S. (2017). A robust deep-learning-based detector for real-time tomato plant diseases and pests recognition. Sensors, 17(9), 2022.
[16] Gonzalez-Huitron, V., León-Borges, J. A., Rodriguez-Mata, A. E., Amabilis-Sosa, L. E., Ramírez-Pereda, B., & Rodriguez, H. (2021). Disease detection in tomato leaves via CNN with lightweight architectures implemented in Raspberry Pi 4. Computers and Electronics in Agriculture, 181, 105951
[17] Guo, Y., Zhang, J., Yin, C., Hu, X., Zou, Y., Xue, Z., & Wang, W. (2020). Plant disease identification based on deep learning algorithms in smart farming. Discrete Dynamics in Nature and Society, 2020.
[18] Harte, E. (2020). Plant disease detection using CNN (Doctoral dissertation, Ph. D. thesis).
[19] Hickey, L. T., Germán, S. E., Pereyra, S. A., Diaz, J. E., Ziems, L. A., Fowler, R. A., ... & Dieters, M. J. (2017). Speed breeding for multiple disease resistance in barley. Euphytica, 213(3), 64.
[20] Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., ... & Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861.
[21] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017). Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708).
[22] Jasim, M. A., & AL-Tuwaijari, J. M. (2020, April). Plant leaf diseases detection and classification using image processing and deep learning techniques. In 2020 International Conference on Computer Science and Software Engineering (CSASE) (pp. 259-265). IEEE.
[23] Liu, B., Zhang, Y., He, D., & Li, Y. (2018). Identification of apple leaf diseases based on deep convolutional neural networks. Symmetry, 10(1), 11.
[24] Mahlein, A. K. (2016). Plant disease detection by imaging sensors–parallels and specific demands for precision agriculture and plant phenotyping. Plant disease, 100(2), 241-251.
[25] Mohanty, S. P., Hughes, D. P., & Salathé, M. (2016). Using deep learning for image-based plant disease detection. Frontiers in plant science, 7, 1419.
[26] Nandhini, S., & Ashokkumar, K. (2021). Improved crossover-based monarch butterfly optimization for tomato leaf disease classification using convolutional neural network. Multimedia Tools and Applications, 80(12), 18583-18610.
[27] Nandhini, S., & Ashokkumar, K. (2021, March). Analysis on Prediction of Plant Leaf diseases using Deep Learning. In 2021 International Conference on Artificial Intelligence and Smart Systems (ICAIS) (pp. 165-169). IEEE.
[28] Rangarajan AK, Purushothaman R, Ramesh A (2018) Tomato crop disease classification using pre-trained deep learning algorithm, In: International Conference on Robotics and Smart Manufacturing (RoSMa2018), Procedia Computer Science 133:1040–1047.
[29] Rangarajan Aravind, K., & Raja, P. (2020). Automated disease classification in (Selected) agricultural crops using transfer learning. Automatika: casopis za automatiku, mjerenje, elektroniku, racunarstvo i komunikacije, 61(2), 260-272.
[30] Reddy, T. V., & Rekha, K. S. (2021, April). Deep Leaf Disease Prediction Framework (DLDPF) with Transfer Learning for Automatic Leaf Disease Detection. In 2021 5th International Conference on Computing Methodologies and Communication (ICCMC) (pp. 1408-1415). IEEE.
[31] Ren, S., He, K., Girshick, R., & Sun, J. (2016). Faster R-CNN: towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6), 1137-1149.
[32] Ren, S., He, K., Girshick, R., Zhang, X., & Sun, J. (2016). Object detection networks on convolutional feature maps. IEEE transactions on pattern analysis and machine intelligence, 39(7), 1476-1481.
[33] Ricciardelli, C., Bianco-Miotto, T., Jindal, S., Butler, L. M., Leung, S., McNeil, C. M., ... & Tilley, W. D. (2018). The magnitude of androgen receptor positivity in breast cancer is critical for reliable prediction of disease outcome. Clinical Cancer Research, 24(10), 2328-2341.
[34] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... & Fei-Fei, L. (2015). Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3), 211-252.
[35] Salimans, T., & Kingma, D. P. (2016). Weight normalization: A simple reparameterization to accelerate training of deep neural networks. arXiv preprint arXiv:1602.07868.
[36] Sandeep, A. S., Krishna, M. V., & Lakshmi, G. J. (2020). TOMATO LEAF DISEASE CLASSIFICATION USING TRANSFER LEARNING. Solid State Technology, 63(1s), 108-115.
[37] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1), 1929-1958.
[38] Subhadra, K., & Kavitha, N. (2021). Smart Agriculture: IoD Based Disease Diagnosis Using WPCA and ANN. Development and Future of Internet of Drones (IoD): Insights, Trends and Road Ahead, 231-249.
[39] Szegedy, C., Ioffe, S., Vanhoucke, V., & Alemi, A. (2017, February). Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 31, No. 1).
[40] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 2818-2826).
[41] Thangaraj, R., Anandamurugan, S., & Kaliappan, V. K. (2021). Automated tomato leaf disease classification using transfer learning-based deep convolution neural network. Journal of Plant Diseases and Protection, 128(1), 73-86.
[42] Tieleman, T., & Hinton, G. (2012). Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn.
[43] Tripicchio, P., Satler, M., Dabisias, G., Ruffaldi, E., & Avizzano, C. A. (2015, July). Towards smart farming and sustainable agriculture with drones. In 2015 International Conference on Intelligent Environments (pp. 140-143). IEEE.
[44] Vallabhajosyula, S., Sistla, V., & Kolli, V. K. K. (2021). Transfer learning-based deep ensemble neural network for plant leaf disease detection. Journal of Plant Diseases and Protection, 1-14.
[45] Van Laarhoven, T. (2017). L2 regularization versus batch and weight normalization. arXiv preprint arXiv:1706.05350.
[46] Zhang K, Wu Q, Liu A, Meng X (2018) Can deep learning identify tomato leaf disease. Adv Multimed 2018:1–10.
[47] Zhang, K., Wu, Q., Liu, A., & Meng, X. (2018). Can deep learning identify tomato leaf disease?. Advances in Multimedia, 2018.
[48] Oltean, M. (2019). Fruits 360 dataset: A dataset of images containing fruits and vegetables.
[49] Tan, M., & Le, Q. (2019, May). Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (pp. 6105-6114). PMLR.
[50] Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. (2018). Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4510-4520).

Claims: 1. A method to detect disease in tomato leaves comprises the steps of:
collecting a dataset of the tomatoes and pre-processing the dataset;
using pre-trained models as a convolutional base for feature extraction of the tomato dataset;
wherein the features extracted are feeding a stack of fully connected layers as classifier followed by hyperparameter tuning of the models, analyzing and comparing the pre-trained models, and obtaining the results based on the compared data.

2. The method as claimed in claim 1, wherein the collected dataset is divided into training data and test data and the pre-processing of the data comprises data normalization and augmentation.

3. The method as claimed in claim 2, wherein the data normalization comprises scaling the pixel values in the dataset images range from 0 to 255 through an image data generator.

4. The method as claimed in claim 2, wherein the data augmentation increases the dataset by introducing diversity and variety in the dataset when there is a limited dataset by introducing various types of transformations and helps to resolve the overfitting problem.

5. The method as claimed in claim 1, wherein the features are extracted using convolution layers and pooling layers from the dataset images.

6. The method as claimed in claim 1, wherein the pre-trained models are selected from Inception3, MobileNet, resnet50, Xception, Densenet121, VGG16, and EfficientNetB0.

7. The method as claimed in claim 1, wherein the layers are frozen before adding the fully-connected classifier using a sequential model.

8. The method as claimed in claim 1, wherein the hyperparameters used in the hyperparameter tuning of the model are model hyperparameters, optimizers hyperparameters, and data hyperparameters.

9. The method as claimed in claim 8, wherein the pre-trained model is trained with the optimizer selected from Adam, SGD, Adagrad, Adadelta and RMSprop.

10. The method as claimed in claim 1, wherein fine-tuning of the pre-trained model is done before the hyperparameter tuning of the models by unfreezing some of the layers of the convolutional base and retraining the network with a lower learning rate and the selected optimizer.