RCNN-based analysis of apple trees leaves for early plant disease detection
Loading...
Supplementary material
Other Title
Authors
Krishnan, Archana
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2024
Supervisors
Ardekani, Iman
Varastehpour, Soheil
Varastehpour, Soheil
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
apple trees
leaf spots
plant disease
modelling
Deformable Convolutional Networks (DCN)
neural networks
Region-based Convolutional Neural Network (RCNN)
image processing
deep learning
leaf spots
plant disease
modelling
Deformable Convolutional Networks (DCN)
neural networks
Region-based Convolutional Neural Network (RCNN)
image processing
deep learning
ANZSRC Field of Research Code (2020)
Citation
Archana, K. (2024). RCNN-based analysis of apple trees leaves for early plant disease detection(Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
https://hdl.handle.net/10652/6744
Abstract
Early identification and accurate localization of plant diseases are essential for enhancing crop output and promoting sustainable agricultural practices. This thesis focuses on creating an innovative hybrid deep learning model that combines Region-Based Convolutional Neural Networks (RCNN) and Deformable Convolutional Networks (DCN) to improve the detection and categorization of illnesses in apple leaves. The hybrid technique seeks to tackle difficulties in precisely identifying regions afflicted by illness, particularly when confronted with complex patterns, uneven shapes, and varying sizes of leaf infections. The proposed structure utilizes the advantages of both RCNN and DCN to address the intricacies of disease patterns, including irregular forms, variable sizes, and diverse textures. Region-Based is utilized to provide region suggestions and extract spatial information, facilitating the detection of possibly sick areas within an image. DCN, due to its ability to cope with deformable patterns, improves feature extraction by collecting complex details sometimes overlooked by traditional convolutional networks. This combination guarantees strong efficacy in detecting and pinpointing disease affected regions. A custom dataset was produced for this study, comprising annotated images of apple leaves with clearly defined diseased areas. Preprocessing methods, such as reducing photos to 100×100 pixels and standardizing data formats, were implemented to guarantee compatibility with the model's input specifications. The dataset was utilized to train the hybrid model, which integrates both feature extraction and classification phases. The framework's principal innovations encompass the incorporation of deformable convolutional layers to manage spatial variability and the application of bounding boxes for the localization of impacted regions. The thesis further examines the technical execution of the hybrid RCNN DCN model, clarifying the model architecture, training setups, and preprocessing pipelines. The hybrid model was executed using TensorFlow and trained for 25 epochs with an 80-20 training-testing split. Evaluation metrics, such as Mean Intersection over Union (IoU), Mean Squared Error (MSE), and accuracy, were employed to assess model performance. The experimental results validated the model's efficacy, achieving a mean Intersection over Union (IoU) of 89.57% and an accuracy of 91.48%, markedly surpassing baseline methods. Visual overlays of predictions on test images validated the model's ability to appropriately pinpoint and classify sick areas.
This guarantees scalability and facilitates customisation to suit various datasets and plant types. Employing comprehensive preprocessing and network optimization methods diminishes computing demands while preserving increased accuracy and dependability. This study highlights the capacity of deep learning to transform plant disease detection through automated, scalable, and precise solutions. The hybrid RCNN-DCN framework establishes a basis for future developments in precision agriculture, underscoring its applicability in real-world applications.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved