IEEE_Conference_Template__2_ (1)

[Virtual Presenter] Building a VGG16 Model for Plant Detection Using Rice Disease Dataset * 2nd Prof.(Dr) Mayuri Metha dept.Computer Enginnering Sarvajanik College of Engineering and Technology Surat, Gujarat mayuri.mehta@scet.ac.in 1st Pooja Patel dept.Computer Enginnering Sarvajanik College of Engineering and Technology Surat, Gujarat poojapatel.comtech23@scet.ac.in 3rdProf.(Dr)Nirali Nanavati dept.Computer Enginnering Sarvajanik College of Engineering and Technology Surat, Gujarat nirali.nanavati@scet.ac.in Abstract—This paper proposes a method for accurately identifying rice diseases from leaf images using the VGG16 convolutional neural network. Preprocessing ensures compatibility with VGG16, while dropout mitigates overfitting. Results demonstrate high training accuracy (0.9881) and validation accuracy (0.6667). The model’s depth captures fine-grained details, while maxpooling reduces complexity. The dataset, encompassing labeled images of rice diseases, enables comprehensive training and rigorous testing. Future work involves dataset expansion and exploring advanced architectures. The success of this approach highlights its potential to revolutionize agricultural practices by facilitating precise disease identification, ultimately enhancing crop management and disease prevention. Index Terms—Convolutional Neural Network (CNN),Plant Disease Identification,Neural Networks. I. INTRODUCTION tomato leaf locations[3].Rice, wheat, and maize are important food grains with many health benefits. They are grown in large quantities and are energy giving food to the world[4].MDFCResNet enhances diagnostic accuracy by leveraging multidimensional data and using a sophisticated compensation algorithm to integrate the recognition results from different dimensions, providing a comprehensive and precise diagnostic tool[5].Large real-world data sets will always include both target and non-target images. If these data sets are not filtered, non-target images will be mixed with target images and mistakenly used for classification.This will prevent the model from training effectively and reduce classification accuracy.[6] The main contributions of the paper are: A new CNN architecture is introduced, combining Inception and Residual connections, which extracts better features and delivers higher performance. The standard convolution is replaced with depthwise separable convolution, significantly reducing the number of parameters without affecting performance. The proposed architecture uses fewer parameters and is faster than other deep learning models. The model’s robustness is tested on three different plant disease datasets: PlantVillage dataset: images taken with a uniform background in a lab setting. Rice disease dataset: images taken in real-time field conditions. Cassava plant dataset: field-conditioned images with multiple leaves. The proposed model outperforms other state-of-the-art deep learning models on all three datasets. The rest of the paper is organized as follows: Section 2 provides the existing literature on the identification of plant diseases using deep learning models. Materials and methods are discussed in Section 3. Section 4 presents the experimental results and discussions. II. LITERATURE SURVEY In the context of India’s agricultural sector, the study of crop diseases reveals critical insights into the economic and social ramifications for the country’s predominantly agrarian population.Recent work on identifying plant diseases based India’s population is approximately 1.44 billion till 2024, making it the most populous country in the world . The agricultural sector remains a significant part of India’s economy, contributing about 18 to the GDP and employing a substantial portion of the workforce. Agriculture is a significant part of India’s GDP. Crop diseases can reduce overall agricultural productivity, leading to a decline in crop outcome and thus negatively affecting GDP. Crop diseases caused by bacteria and fungi significantly reduce crop yield and quality.Detecting disease symptoms early is challenging, especially in large farms. In developing countries, experts visually identify diseases, but this method is slow and expensive. Using smart devices for automatic disease identification offers a promising solution, reducing costs and improving efficiency[1]. Diseased plants often display noticeable marks or lesions on leaves, flowers, or fruits. Each disease or pest condition typically has a distinct visible pattern, aiding in diagnosis. Leaves are commonly the first place to observe symptoms, with many diseases manifesting there initially[2]. The Faster RCNN algorithm to detect diseased tomato leaves, which can both recognize tomato diseases and detect.

[Audio] array. The dataset and labels for all images are then separated. The model is trained on a dataset consisting of images of various diseased plant leaves. The labeled data is stored in rice disease files, which are accessed during the training phase. The model includes convolution layers followed by max-pooling layers, and 0.25 of the data is dropped out to prevent overfitting. The output is then flattened and fed into a dense network. The final layer uses softmax activation to predict the disease from the leaf image. The Adam optimizer is used to minimize the loss function. The framework automatically identifies and preprocesses the leaf image for prediction, generating 15 probability values for 15 labels. The label with the highest probability is predicted as the disease or result for that image. Fig. 1. Sample Image of Rice Disease . on deep learning models. SK MAHMUDUL HASSAN AND ARNAB KUMAR MAJI at [1]proposed a novel CNN model based on the inception and residual connection that can effectively classify the diseases in plants. Pre-trained VGG model with inception layer termed asINC-VGGN to identify the different corn and rice diseases. They replaced the fully connected layers of VGGNet and added two inception layers. The average testing accuracy obtained in rice diseases is 0.92 and 0.8038 in corn diseases. Shallow CNN from the pre-trained VGG16 model to identify different diseases in corn, apple, and grape[2]. They have used only the first four-layer of VGG16 architecture, a global pooling layer. Shallow CNN was used to extract the features and PCA to reduce the dimensions. They obtained an f1-score of 0.94 using SVM and RF respectively. Self-attention CNN was used to identify different crop diseases. Attention network is effective in extracting the image features from the critical region. The review in [4] mainly focuses on recent advances in using machine learning and deep learning for diagnosing plant diseases and how these techniques speed up disease detection. The study explains the importance of detecting pests and diseases in plants, the different datasets used for this purpose, the performance measures for evaluating these models, the challenges in this field, and potential future improvements. Early crop disease identification methods rely heavily on the experience of farmers or experts, leading to subjective awareness, low recognition efficiency, and high error rates. The advent of image processing technology improved identification by converting images to digital signals for processing, offering better reproducibility and accuracy. However, it falls short for complex disease identification. Deep learning addresses these challenges by capturing and analyzing photos of diseased crops with neural networks, offering faster and more accurate identification. Recent advancements combine IoT with deep learning, using real-time data from sensors to enhance disease detection and provide timely information on smart devices[5]. 1) CONVOLUTION NEURAL NETWORK: In machine learning, CNNs use a different approach to regularization that is simpler than traditional regularization methods. Here's a breakdown of the layers involved A.Input Layer In the input layer, the model receives the initial data. At this stage, the number of neurons matches the number of features. For an image, this means each pixel is treated as a separate feature. The input data is split into two sets: one for training the model and the other for testing it. Typically, the larger portion of the data is allocated for training, while a smaller portion is reserved for testing. B. Hidden Layer This layer receives III..

[Audio] improvement in disease detection accuracy, making it a valuable tool for early disease diagnosis and effective crop management in the agricultural sector. C. IMAGE PREPOSSESSING 2) ARCHITECTURE OF THE VGG16 DEPP CNN MODEL: A convolutional neural network (ConvNet) is an artificial neural network designed for image processing, comprising an input layer, an output layer, and multiple hidden layers. VGG16, a notable ConvNet, excels in computer vision tasks due to its depth and use of small 3x3 convolution filters. With 16 weight layers and about 138 million trainable parameters, VGG16 effectively captures detailed features in images, making it highly suitable for applications like plant disease detection in agriculture. We employed the VGG16 convolutional neural network Image preprocessing is a pivotal stage in readying images for the VGG16 model, a deep learning architecture renowned for its effectiveness in image recognition tasks. Primarily, this process ensures that input images adhere to the specific requirements of the VGG16 model, thereby optimizing its performance. Initially, the image is loaded from the file system using libraries like PIL or Keras's image module. Subsequently, the image is resized to dimensions of 224x224 pixels, aligning with the input size expected by the VGG16 model. This resizing step is critical for maintaining compatibility and consistency in model input. Following this, the resized image is transformed into a numerical array , facilitating processing by the deep learning model. Additionally, an extra dimension is appended to represent the batch size, essential for models expecting batches of images. Once transformed into an array or tensor, the pixel values of the image undergo normalization, aligning them with the distribution expected by the VGG16 model. Typically, this involves scaling pixel values to a specific range and normalizing based on the mean and standard deviation of the ImageNet dataset, on which the VGG16 model was trained. Through adherence to these preprocessing steps, images are suitably prepared for input into the VGG16 model, ensuring compatibility and facilitating accurate and reliable predictions. This meticulous preprocessing ensures that input images possess the requisite characteristics akin to the training data, thereby optimizing the performance of the VGG16 model in various image recognition tasks. Fig. 3. Proposed VGG16 approach in identification of plant diseases. IV. EXPERIMENTAL RESULTS to evaluate the performance of the model,we consider different performance statistics such as number of parameters, accuracy. Table shows the performance of the implementation model on rice disease detection. Table.II Parameter required in VGG-16 Model. after 60 epochs of training, In the rice dataset, the model achieves the highest training accuracy of 96.43, training loss of 0.0806, validation accuracy of 99.66, and validation loss of 0.6667. Fig. 4. Proposed VGG16 approach in identification of plant diseases. (CNN) architecture to develop a robust plant disease detection model using a rice disease dataset. VGG16, known for its deep architecture with 16 weight layers, has been widely recognized for its effectiveness in image classification tasks. The model's architecture consists of multiple convolutional layers with small receptive fields, followed by max-pooling layers, which contribute to its strong feature extraction capabilities. By leveraging the pre-trained VGG16 model on ImageNet and fine-tuning it with kaggle rice disease dataset, we aimed to achieve high accuracy in identifying various rice diseases from leaf images. The dataset comprised labeled images of rice plants affected by common diseases such as rice blast,.

[Audio] Layer Output Shape Parameters conv2d None, 224, 224, 16 432 max pooling2d None, 74, 74, 16 0 conv2d 1 None, 74, 74, 16 4608 max pooling2d 1 None, 74, 74, 16 0 diseases, including their classifications, types, and the quantity of available image data for each disease, which is crucial for studying and developing methods for disease detection and management in rice cultivation. conv2d 2 None, 74, 74, 16 18432 max pooling2d 2 None, 8, 8, 64 0 V. DISCUSSION conv2d 3 None, 8, 8, 128 73728 max pooling2d 3 None, 4, 4, 128 0 conv2d 4 None, 4, 4, 256 294912 max pooling2d 4 None, 2, 2, 256 0 dropout None, 2, 2, 256 0 flatten None, 1024 0 dropout 1 None, 1024 0 dense None, 250 256250 dense 1 None, 100 25100 dense 2 None, 3 303 multicolumn3—r—Total Parameters 673765 TABLE II PARAMETERS REQUIRED IN VGG-16 MODEL After splitting the dataset in to 70-30 taiining and validation set we train the model upto 60 epoch and evaluate the performance on training and validation set. The table below provides detailed information about the rice disease dataset, including the class, plant name, disease name, type of disease, and the number of images available for each disease type. The proposed method employs VGG16 CNN to classify rice diseases from leaf images, leveraging its effectiveness. Image preprocessing, including resizing and normalization, ensures compatibility with VGG16. The model's depth enables capturing fine-grained details, crucial for disease identification. Max-pooling layers reduce computational complexity while retaining essential information. The dataset, comprising labeled images of various rice diseases, ensures comprehensive training. Splitting the dataset for training and validation rigorously tests the model's generalization capability. Training with the Adam optimizer achieves high accuracy and effective generalization. The dropout rate in dense layers mitigates overfitting, ensuring model robustness. The detailed parameter count highlights the model's complexity. The high validation accuracy demonstrates proficiency in predicting unseen images. Future work could involve expanding the dataset and exploring advanced architectures for further improvements. The success of the VGG16based model underscores its potential in revolutionizing agricultural practices. TABLE III DATA DESCRIPTION OF RICE DISEASE DATA VI. CONCLUSION Class Plant Name Disease Name Types of Disease No of Image C1 Rice Bacterial blight Bacterial 1584 C2 Rice Blast Fungal 1440 C3 Rice Brown Spot Fungal 1600 The proposed method for plant disease detection using VGG16 demonstrates significant potential in accurately identifying rice diseases from leaf images. The high accuracy achieved, coupled with effective preprocessing and model architecture, underscores the viability of this approach in practical agricultural applications. Future work could involve expanding the dataset with more diverse samples, exploring additional CNN architectures, and developing real-time diagnostic tools to further enhance crop management and disease prevention. REFERENCES [1] Hassan, Sk Mahmudul, and Arnab Kumar Maji. "Plant disease identification using a novel convolutional neural network." IEEE Access 10 (2022): 5390-5401. [2] Li, Lili, Shujuan Zhang, and Bin Wang. "Plant disease detection and classification by deep learning—a review." IEEE Access 9 (2021): 56683-56698. [3] Zhang, Yang, Chenglong Song, and Dongwen Zhang. "Deep learningbased object detection improvement for tomato disease." IEEE access 8 (2020): 56607-56614 [4] Joseph, Diana Susan, Pranav M. Pawar, and Kaustubh Chakradeo. "Realtime Plant Disease Dataset Development and Detection of Plant Disease Using Deep Learning." IEEE Access (2024). Fig. 5. training and validation accuracy in rice datasets.. [5] Hu, Wei-Jian, Jie Fan, Yong-Xing Du, Bao-Shan Li, Naixue Xiong, and Ernst Bekkering..

[Audio] [7] Shrestha, Garima, Majolica Das, and Naiwrita Dey. "Plant disease detection using CNN." In 2020 IEEE applied signal processing conference (ASPCON), pp. 109-113. IEEE, 2020.