Exploring Neural Network Efficacy: Comparative Analysis of Transfer Learning Approaches in Tomato Leaf Disease Detection
Year: 2023
Type: dissertation
Abstract: Image classification has come a long way over the past few decades.With the help of deep learning networks, we saw their applicability in our daily lives to advance safety-critical systems.This thesis advances machine learning and deep learning for image classification by structuring a baseline convolutional neural network (CNN) model and comparing it to four pre-trained models to identify tomato leaf diseases.These models are DenseNet201, InceptionV3, NasNetLarge, and MobileNetV2.The aim is to show that transfer learning can improve a CNN model's performance beyond its baseline capabilities and to compare the four pre-trained models' performance in this specialized task.A baseline CNN model was initially constructed following a deep learning standard architecture to analyze the performance of the dataset of tomato leaves containing 10,000 images.The model's evolved accuracy measures showed its adaptation to the dataset, achieving 80.20% training accuracy and 85.50% test accuracy.After developing the baseline model, the research strategically refined the four pre-trained models.Fine-tuned models produced remarkable and illuminating results.DenseNet201's adaptability led to 99% accuracy on the test dataset after fine-tuning, which showed the model's excellent precision and adaptability.In addition to DenseNet201, after fine-tuning, InceptionV3, NASNetLarge, and MobileNetV2 significantly improved their classification abilities.These improvements were measured in accuracy, classspecific precision, recall, and F1 scores to show each model's disease categorization strengths and weaknesses.These results show how transfer learning transforms CNNs, especially for specialized image classification applications.Comparing the baseline model to the fine-tuned pre-trained models shows how transfer learning improves model performance.The in-depth analysis of pretrained models shows how different architectures and training procedures affect the result of model accuracy, highlighting the complex relationship between model design and task specificity.This discovery opens the door to profound learning advances for complex and specialized image classification problems.The findings stimulate further study on machine learning in image classification and beyond.
Author: Nayeem, Mahedi
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7215386849