Description:Field of Invention
Agriculture holds a crucial position in the global economy, serving as the main source of food, fibre, fuel, timber, and more, thereby ensuring socio-economic stability. The tomato ranks among the most widely cultivated vegetable plants globally. Initially from Peru in South America, its cultivation has expanded worldwide, occurring in both fields and greenhouses. The states known for tomato production include Andhra Pradesh, Karnataka, Maharashtra, Uttar Pradesh, Orissa, Assam, Madhya Pradesh, and Bihar. The productivity of tomatoes is enhanced due to the advantages associated with their cultivation; however, the full potential of production remains untapped due to the crop's vulnerability to various pests and diseases, leading to reduced yields. Diseases significantly impact production, leading to substantial losses in the agricultural economy. The onset of the disease is instigated by various pathogens, including viruses, bacteria, nematodes, and fungi. The prevalent diseases affecting tomato plants include Bacterial Speck, Bacterial canker, Grey leaf spot, Tomato spotted wilt virus, Powdery mildew, Grey mould, Early blight, Late blight, Bacterial spot, Leaf mould, Mosaic virus, Septoria leaf spot, two-spotted spider mites, Yellow leaf curl virus, and Target spot.
Background of the Invention
Rapid identification of plant diseases is of the utmost importance in the agricultural sector. Since the original structure of leaves changes for many diseases, detecting them through leaves is a commonly utilised way for figuring out the condition. An expert's keen vision and deep understanding of the factors that cause crop diseases are required for disease identification by naked eye. Most recently, there has been a chance to increase the market for computer vision applications in agriculture and broaden the practice of precision plant protection thanks to advancements in computer vision. To simplify the data and make patterns more apparent to learning algorithms, a domain expert is typically required to identify the majority of the applied features in classical machine learning techniques. There have been numerous uses of Machine Learning (ML) models for plant disease detection and categorisation. Deep Learning (DL) is an advanced subfield of machine learning that use artificial neural networks with a hierarchical structure to do machine learning. Deep learning involves training a computer model to identify objects in photos. To train their models, deep learning systems use massive amounts of labelled data in conjunction with neural network topologies that can learn features automatically from the data. Several visualisation techniques are used in conjunction with many DL architectures that have been designed or customised in order to identify and categorise symptoms of plant diseases. Deep Learning algorithms aim to incrementally learn high-level features from data, which is their most significant advantage. Because of this, domain knowledge and hard core feature extraction are no longer necessary (US9968045B2).
Tomato plant disease detection sometimes begins with the crop leaves, which often serve as symptoms for more serious diseases. Therefore, in order to draw conclusions from the prior study, a comprehensive literature review was conducted. To determine if tomato leaves have diseases like early blight or powdery mildew, the SVM approach is employed. To classify diseases in tomatoes, the approach uses support vector machines (SVMs) after collecting leaf information using the Gabor wavelet transform methodology. As an added bonus, disease identification has been approached using three distinct kernel types: Invmult, Cauchy, and Laplacian. By utilising three distinct classifiers—ANFIS, FIS, and MLBPNN—conventional image processing methods were employed to convert RGB images into a CIE XYZ colour space model.
Brahim showcased two methods for identifying ten different types of tomato leaf diseases. The first method involves starting from scratch when training CNN models based on GoogLeNet and AlexNet. Second, we have the optimisation methods of Random Forest (RF) and Support Vector Machine (SVM) that make use of transfer learning. According on the findings of the experiments, the pre-trained CNN model outperforms the RF and SVM classifiers. The six different tomato illnesses were identified using photos of the leaves by Manpreet Kaur and Rekha Bhatia using a pre-trained convolutional neural network (CNN) model called ResNet101. The experiment utilises tomato leaf photos taken from the PlantVillage collection. Early blight, late blight, and leaf mould are the three disease types employed in the experiment, which is conducted using the PlantVillage dataset. Use of the CNN-learned features at various processing levels is crucial to the proposed model.
An automated learning architecture known as "Deep Learning" developed from multi-layer artificial neural networks. In order to divide data into supervised, unsupervised, and reinforcement learning categories, DL architectures are employed to create information ranging from simple to complicated. Numerous picture recognition difficulties have been solved and study areas such as medical diagnosis, natural language, and automatic plant disease detection have seen greater success thanks to applications incorporating deep learning models. A novel method is the use of deep learning for the identification of plant species and leaf diseases (CN109022315B).
Diseases can cause a wide variety of problems for plants. There are a lot of similarities between plant families and a lot of differences within them when it comes to background, occlusion, position, colour, and lighting, which makes plant disease classification difficult. Many people believe that diseases are one of the main obstacles to growing tomatoes. Diseases wreak havoc on tomato crops, cutting into yield and wreaking havoc on the agricultural sector. The quantity and quality of tomato crops can be greatly enhanced by taking measures to protect them from illnesses. Therefore, it is helpful to provide precise disease identification in order to choose treatment against widespread harms. As an added complication, farmers have the unpleasant chore of constantly checking tomato plants for signs of infection. Therefore, a lot of work has gone into developing a method that uses photographs of leaves to automate the illness classification process. One DL approach is the Deep Convolution Neural Network (DCNN), which uses several layers of neural networks. DCNN models operate under the premise that deeper networks, equipped with a greater variety of nonlinear mappings and richer feature hierarchies, can more accurately approximate the target function. Multiple feature extraction phases allow DCNN to automatically learn representations from data, which is the primary reason for its remarkable learning ability. The multi-layered, hierarchical structure of DCNN, offers it the ability to extract low, mid, and high-level properties of the tomato leaves. Features at lower and mid-levels are combined to form features at a higher level, which are more abstract.
Summary of the Invention
The most devastating aspect of agriculture is plant diseases, which drastically lower crop yields and lead to financial losses. As a result, the agricultural sector greatly prefers the early and correct diagnosis of these diseases. It gets difficult when diseases spread over crops and production drops. Because of its exceptional performance, Deep Learning (DL) has become the preferred method for identifying plant diseases. The suggested study uses a model called MX-MLF2, which stands for Modified-Xception based Multi-Level Feature Fusion. Classification of tomato leaf diseases is achieved by the use of feature fusion and multi-level feature extraction. The accuracy of the tomato leaf disease predictions is further improved by using a fine-tuning and transfer learning strategy.
Brief Description of Drawings
Figure 1: Multi layer feature fusion Framework.
Figure 2: Architecture of the Enhanced xception Model
Detailed Description of the Invention
Deep Convolutional Neural Networks (DCNNs) have recently emerged as a strong tool for feature extraction and data transformation, automatically bridging the semantic gap through hierarchical picture abstraction and transforming raw data into high-level representation. At the beginning, scientists had to start from square one when training a DCNN model. It is challenging to obtain improved accuracy without a large amount of labelled data, which increases the burden of overfitting and the associated costs in terms of both time and computer resources. In order to train the model with insufficient data, transfer learning and fine-tuning were the most effective methods. By first training the model on a larger dataset, transfer learning is able to apply what it has learnt to a smaller dataset. Typically trained on ImageNet, a model can be fine-tuned by training it using the existing parameters. These benefits have contributed to the fine-tuning process's meteoric rise in popularity. While using a fine-tuning method, most of the works ignored extra layers that included useful semantic information and instead concentrated on the feature maps from the last convolution layer. Multi-Layer Feature Fusion (MLFF), depicted in figure 1, is suggested as an improvement to feature extraction, fine-tuning technique, and integration of multi-layer features that is based on the premise that adding more layers can be beneficial. By delving into various sections of DCNN, MLFF can unearth information that the routine structure of DCNN has missed, thereby improving both local and global features.
Features can be extracted from pre-trained models' intermediate layers using the feature fusion method. In order to improve the classifier model's feature capabilities, the extracted features are joined or concatenated to retrieve multi-level information from input photos. One feature map is produced by each DCNN layer. Despite this, the advantages of extra layers are mostly ignored in the literature, which primarily deals with the feature maps of the final convolution layer. There may be feature discrimination capability in the feature information that is concealed in different levels. Hence, the paper suggests a multi-layer feature fusion method that uses feature maps from many layers rather than just the last convolution layer. Then, just like a regular CNN, the fused features are sent to the next layer or classifier. One method of using MLFF that improves the global feature information is the interlayer fusion. During CNN forwarding propagation, pooling causes certain feature information to be lost. By combining global features and increasing the feature diversity of final feature maps, multi-layer feature fusion incorporates the feature information of earlier layers into the fusion feature map. This process is known as interlayer fusion. As a result, it is able to return and compensate for the lost features. Subsequently, the classifier receives the outcome of the feature map integration performed by the intern layer fusion module.
The process of reusing a model that has been trained on a predictive modelling problem that is similar is referred to as transfer learning. To apply a model that was trained on one problem to another related problem is the general idea behind transfer learning. Both the weight initialisation scheme and the feature extraction approach can be employed by transfer learning to expedite the training of neural networks. One method used in deep learning is transfer learning, which entails training a neural network model on a problem that is comparable to the one being solved. A new model is trained on the problem of interest using one or more layers from the trained model. Reduced generalisation error and shorter neural network model training times are two advantages of transfer learning. The two most common ways to use transfer learning are feature extraction and weight initialisation. It is possible to modify the weights of previously trained layers to fit the current problem by using them as a starting point. Another possibility is that the network's weight has not been adjusted to account for the new issue. It is possible to train new layers to understand the output after the reused layers. For the sake of this application, transfer learning is considered a feature extraction scheme. Some variants of these applications may require starting with a different problem and not training the model's weights on it. Afterwards, honing all of the learnt model's weights using a modest learning rate.
The Xception architecture, which was presented by Francois Chollet, is an expansion of the Inception architecture. The Inception design is a linear stack of depthwise separable convolution layers with residual connections. The goal of the depthwise separable convolution is to decrease the memory and computational demands. With the exception of the first and last modules, all fourteen of Xception's 36 convolutional layers feature linear residual connections. In Xception, channel-wise and space-wise feature learning are separated by the separable convolution. By shortening the sequential network, the residual connection aids in resolving representational bottlenecks and vanishing gradients. Instead of concatenating the outputs of several layers, the shortcut connection makes them available as input to the later layer through a summing procedure. The Xception model, which uses depth-wise convolutions, served as an inspiration for the proposed MX-MLF2 model. The suggested model extracts features from several levels by modifying the Xception's original architecture. The inception module, depthwise separable convolution layers, and residual blocks are its three main parts. In order to enhance the illness classification performance, the features are combined once they have been extracted from the intermediate convolution layers. The features are taken and inputted into the Global Average Pooling (GAP) from the MX-MLF2 model's intermediate layers. By lowering the number of parameters in the model, the GAP layer lessens the likelihood of overfitting. In order to reduce the tensor's dimensions from h x w x d to 1 x 1 x d, the GAP layer employs dimensionality reduction. The GAP layer, which stands in for the fully linked layers utilised by traditional CNN, is responsible for generating output for each feature map and calculating the average of all of its pixels. After that, the GAP uses these output features to create a feature vector. For classification, the softmax layer receives a combined feature vector from all of the GAP layers. In addition, a tomato dataset is used to train and fine-tune the last few layers through the parameter transfer. Faster convergence and improved generalisation are the results of fine-tuning a network with transfer learning rather than building a network from scratch with randomly initialised weights. Model training makes use of the tomato dataset to fine-tune the classification layers' weights, while the pre-trained Xception model is used to initialise the convolution layers' weights. The MX-MLF2 model incorporates a new softmax layer in place of the old one. The softmax layer takes the target class's highest predicted value and uses it to do class prediction. There are exactly as many modified classes in the softmax layer as there are tomato classes overall. As shown in Figure 2, the MX-MLF2 model is conceptualised for the purpose of disease categorisation in tomato leaves.
One idea in TL is fine-tuning (FT), which involves freezing the initial layers of a pre-trained model. To improve the model's performance, the weights of the final layers are adjusted. The FT procedure involves updating the weight values of the MX-MLF2 classification layer and retraining it with the dataset of images of tomato leaves. In addition, the Adam optimiser and a learning rate value of 0.001 are used to conduct the model's FT. In a multi-class classification problem, the softmax function is utilised in a fully connected layer. Using the information acquired from the prior layers, the class prediction is carried out by the softmax function. The layer takes into account the possibilities of each class and selects the one with the highest likelihood. The main goal is to use the TL idea to combine features from the intermediate layers and fine-tune MX-MLF2's classification layer. The pre-trained model is fine-tuned using the tomato dataset by employing the TL principle. There is a limit of 14 epochs per model in the experiment. , Claims:The scope of the invention is defined by the following claims:

Claim:
1. A System/Method to Detect Tomato Leaf Disease using an Enhanced xception model Deep Convolutional Neural Networks comprising the steps of
a) Multi-level Features are extracted from the tomato leaf images by combining the features from different layers at a time.
b) From the extracted multi-level features select the relevant features which are best suitable to train the model for classification.
c) Train the model with selected relevant features and the classification will be performed.
2. According to claim 1, Deep Convolutional neural network is used to extract the multi-level features from the given set of images.
3. According to claim 1, for selecting the relevant features for training the model, transfer learning is used.
4. According t claim 1, for classification an enhanced Xception model is used to classify the tomato leaf diseases.