Lightweight CNN-RNN model for tomato leaf disease detection
Loading...
Supplementary material
Other Title
Authors
Le, An Thanh
Author ORCID Profiles (clickable)
Degree
Master of Computing
Grantor
Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
Date
2024
Supervisors
Ardekani, Iman
Shakiba, Masoud
Shakiba, Masoud
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
tomato plants
leaf spots
image processing
neural networks
modelling
Internet of Things (IoT)
leaf spots
image processing
neural networks
modelling
Internet of Things (IoT)
ANZSRC Field of Research Code (2020)
Citation
Le, A.T.. (2024). Lightweight CNN-RNN model for tomato leaf disease detection (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Computing) Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology https://hdl.handle.net/10652/6441
Abstract
This research employs a hybrid lightweight model combining Convolutional Neural Networks and Recurrent Neural Networks to detect tomato plant diseases from leaf images. Traditional image classification models often require substantial computational resources to achieve high accuracy, limiting their practical application. The primary goal of this study is to develop a lightweight model that can be easily implemented on low-cost Internet of Things devices while maintaining high accuracy with real-world images, thereby making tomato disease detection more practical for real-world use. Additionally, this research aims to deepen the understanding of hybrid Convolutional Neural Networks - Recurrent Neural Networks models, investigate the applicability of Liquid Time-Constant networks - or liquid neural networks - in hybrid techniques, and enhance model generalization through augmentation techniques that mimic real-life conditions.
The methodology leverages the strengths of Convolutional Neural Networks for extracting high-level image features and Recurrent Neural Networks for capturing temporal relationships, thereby improving model performance and accuracy. The proposed model incorporates Closed-form Continuous-time Neural Network, a lightweight variant of Liquid Time-Constant networks, which was recently developed to effectively capture complex temporal patterns with greater expressivity than other Recurrent Neural Networks. The integration of the Neural Circuit Policy helps capture long-term dependencies in image patterns while maintaining stability and sensitivity to short-term causality, thereby reducing overfitting. Additionally, random rotation and brightness and contrast adjustments are applied to the training data to reflect real-world conditions, further enhancing model generalization.
The results show that the hybrid models outperform their single Convolutional Neural Networks counterparts in both accuracy and computational cost. The designed model utilizing Closed-form Continuous-time with Neural Circuit Policy achieved the highest performance with 97.15% accuracy on the test set, compared to around 94% for pre-trained Convolutional Neural Networks models. The designed model also demonstrated nearly the lowest computational cost among the models tested. Furthermore, training the models on augmented data improved their accuracy and generalization capabilities.
In conclusion, this research provides solid evidence that the hybrid Convolutional Neural Networks - Recurrent Neural Networks model is an effective approach to improving accuracy without increasing computational cost. The study also highlights the enhanced performance of using liquid neural networks in this hybrid approach, an area not extensively explored in previous research, thereby expanding the application value of liquid neural networks. This approach has broader implications, suggesting that similar methods could be applied and validated in other fields, such as healthcare, manufacturing, and environmental monitoring.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved