Enhancenet: A Comprehensive Approach to Scaling Convolutional Neural Network

Abstract

Early detection of plant diseases is vital for reducing crop losses and ensuring sustainable agriculture. This study presents a novel approach using a Compound Scaling Convolutional Neural Network (CS-CNN) optimized with Bayesian Optimization (BO) to identify diseases in tomato and chitrak plants, which are highly susceptible to pathogens affecting yield and quality. Traditional methods, such as manual inspections and simple algorithms, are often labour-intensive and prone to errors. The proposed CS-CNN enhances disease detection by leveraging compound scaling, which adjusts depth, width, and resolution simultaneously to better recognize complex patterns in leaf images. Bayesian Optimization fine-tunes critical hyper parameters, such as learning rate and filter size, to improve model accuracy and reduce overfitting. Experiments on a dataset of tomato and chitrak leaves demonstrate that the BO-enhanced CS-CNN outperforms conventional CNNs in classification accuracy, computational efficiency, and robustness. This method offers a scalable, efficient, and reliable solution for real-time plant health monitoring and timely disease diagnosis.

Introduction: The Chitrak plant (Plumbago zeylanica) is valued for its medicinal properties, but diseases caused by bacterial, viral, or fungal infections can impact its health and yield. Traditional disease detection methods, such as manual inspection and biochemical testing, are often time-consuming and inaccurate. Convolutional Neural Networks (CNNs) offer an efficient, automated alternative, but their performance depends on architecture and dataset quality. Compound scaling, introduced in EfficientNet, optimizes CNNs by simultaneously adjusting input resolution, width, and depth, enhancing accuracy without excessive computational costs. This study employs Compound Scaling CNNs (CS-CNNs) to classify plant diseases using annotated images, ensuring high accuracy while enabling deployment on resource-constrained devices like smartphones. This approach supports sustainable farming by providing an accessible, real-time disease detection tool, helping preserve the medicinal value of Chitrak plants.

Objectives: This study aims to develop an efficient disease detection system for Plumbago zeylanica (Chitrak), a medicinal plant prone to infections. Traditional diagnostic methods are slow and inaccurate, making automation essential. Convolutional Neural Networks (CNNs) offer a solution, but their accuracy depends on architecture and data quality. To enhance performance efficiently, this study utilizes Compound Scaling CNNs (CS-CNNs), optimizing input resolution, width, and depth. Designed for resource-limited devices like smartphones, the model enables real-time disease detection, supporting sustainable farming and preserving the medicinal value of Chitrak plants.

Methods: This study develops a CNN-based framework in PyTorch to classify Plumbago zeylanica (Chitrak) leaf health using a dataset of 500 images. Focal loss addresses class imbalance, while stochastic depth improvements, including adaptive survival probability and learning rate scheduling, enhance regularization. EfficientNet with Macon blocks extracts features, followed by global average pooling and a fully connected layer for classification. Optimized for resource-limited devices, the model enables real-time disease detection for sustainable farming.

Results: The model achieved impressive results with 95% accuracy in classifying healthy leaves and 99% accuracy for unhealthy leaves, with only six misclassifications out of 200 samples. Training performance showed a significant reduction in loss and a sharp increase in accuracy, from 55% to over 95%, indicating effective learning but suggesting potential overfitting. The model demonstrated high precision and recall, successfully balancing accurate disease detection and minimizing false positives. Data augmentation improved the model's ability to handle variations in leaf orientation, size, and lighting, ensuring strong generalization to unseen data.

Conclusions: In this study, we explored the application of Convolutional Neural Networks (CNNs) with compound scaling for the task of identifying healthy and unhealthy leaves from Chitrak images. By utilizing compound scaling, we effectively designed a CNN architecture that balances depth, width, and resolution, allowing the model to efficiently process and classify images with diverse features and condition.