Aim:
To address the significant challenge of identifying plant diseases using deep Residual Network.
Synopsis:
The proposed work addresses the challenges of identifying plant diseases using a lightweight Deep Residual Network. The primary objective is to develop an efficient and accurate leaf disease identification system. The challenges in the plant disease identification include the need for robust models capable of distinguishing subtle visual cues associated with various disease as well as the requirement for large and diverse datasets for effective training. The proposed method begins by visualizing sample images from dataset, offering insight into diverse plant diseases under consideration. The code creates an image data store, explores dataset characteristics and splits it into training and validation sets. To address the challenge of limited data, data augmentation techniques such as random rotation, reflection and translation are applied to artificially increase the dataset diversity. The proposed work employs a pre-trained ResNet50 model as the foundation and customizes it by replacing fully connected layers with tailored architecture.
Proposed System:
In the proposed methodology for leaf disease detection based on a lightweight deep residual network. The process begins with an exploration of the dataset containing plant images, enabling a thorough understanding of class distribution and sample characteristics, subsequently, an imageDatastore is established, incorporating subfolders for each class, facilitating efficient dataset management. The dataset is then divided into training and validation sets with 80% designed for training and 20% for validation. Data augmentation techniques including random rotation, reflection and translation are applied to enhance the diversity of the training samples. Leveraging a pretrained ResNet-50 model, its fully connected layers are replaced with custom layers such as batch normalization, dropout and additional fully connected layers, tailored for the leaf disease classification task. The modified layers are connected to construct a new layer graph and crucial layers responsible for feature extraction are frozen to retain learned features during training. Training configurations, including the choice of the Adam optimizer, initial learning rate, mini batch size, and training duration are established. The model undergoes training using the augmented dataset with the training progress visualized through plots. Model evaluation is conducted using a separate validation dataset and performance metric such as accuracy are computed. The trained model is then employed to predict classes for sample images in the validation set and a confusion chart is generated to visualize the model classification performance
Reviews
There are no reviews yet.