Abstract:
Acquiring comprehensive and varied datasets for training Machine Learning (ML) models
poses a significant worldwide challenge, especially in areas that need intricate data,
such as detecting plant diseases. This thesis introduces GreenCrowd, an innovative
framework designed to address this challenge through crowdsourcing. This framework
not only facilitates the collection of authentic data but also rewards contributors, thereby
incentivizing participation and enhancing data quality. The project features two core
components: a crowdsourcing mobile application and a sophisticated grape disease detection
model. The mobile application, developed with Flutter, enables users primarily
farmers to upload images of grape leaves along with descriptions of observed diseases.
These images are stored in a centralized database with placeholder values for attributes
such as image quality and plant verification, managed via a Flask API. The image processing
pipeline in GreenCrowd involves several stages. Initially, a basic green color
detection algorithm determines the presence of grape plants. Images confirmed to contain
grape leaves are then evaluated for quality using an Image Quality Assessment
(IQA) model. The IQA model assigns scores to the images, and a Q-learning algorithm
ranks them on a scale from 1 to 5. Images that meet a predefined quality threshold are
used to enhance the disease detection model through transfer learning, incorporating
new disease data into the model continuously. This framework creates dynamic feedback
loop, where the dataset’s quality improves through community participation, and
the detection model’s accuracy is refined over time. The effectiveness of Green Crowd
has been demonstrated in the context of grape leaf disease detection, highlighting its
impact on both data collection and practical disease identification for end-users.
