Rodrigo, Kayne Uriel L.; Marcial, Jerriane Hillary Heart S.
An enhancement of convolutional neural network algorithm applied in rice leaf disease classification mobile application
- Undergraduate Thesis: (Bachelor of Science in Computer Science) - Pamantasan ng Lungsod ng Maynila, 2025
ABSTRACT: This study focuses on enhancing a rice leaf disease image classification algorithm to address its limitations, such as inconsistent image segmentation processes, lack of global context understanding, parameter-heavy computer vision models, and unoptimized hyperparameter values. Hence, the researchers provided three contributions. First, the researchers integrated a U2-Net salient object detection algorithm for an intelligent image background removal process into the rice leaf disease dataset. It reaches a consistent object extraction and reaches an optimal level of mean IOU of 86%, a 3% increase compared to the traditional method. Second, the researchers integrated changes to the architecture of the traditional convolutional neural network algorithm by implementing MobileViTV2_050. It offers a lightweight hybrid CNN + Vision Transformer architecture. This resulted in a 6% increase in Top-1 Accuracy from a baseline of 81%, reduced 13.1 million parameters from the traditional algorithm, and having an output model that is significantly 163.62 MB lesser from a baseline of 168 MB. This makes it susceptible to a lightweight level, suitable for mobile devices. It can also accept pre-trained weights, which can potentially increase its accuracy up to an overall 99% accuracy. Lastly, the researchers also tested GridSearchCV with 3-fold cross-validation, which is tested on both traditional CNN and the proposed MobileViTv2_050, resulting in a 2% and 3% F1-Score increase. Overall, the enhanced algorithm became applicable for low-end devices to accommo