PyCon APAC 2025

Presenter: Kayne Uriel K. Rodrigo
School: Pamantasan ng Lungsod ng Maynila

Rice is one of the staple foods across the Asia-Pacific region. Integrating new trends in artificial intelligence will help secure rice production through the future years.
In this talk, I will present an enhancement to rice leaf disease image classification using transfer learning with the MobileViTV2 model. Traditional Convolutional Neural Networks (CNNs) have been the backbone of many image classification tasks, but their resource-intensive nature makes them challenging for mobile deployment.

We employed MobileViTV2_050, a lightweight model combining CNN’s local feature extraction with Vision Transformers’ global context learning. This model uses a separable self-attention mechanism to improve classification performance while significantly reducing computational load.

Attendees will learn how transfer learning with MobileViTV2_050 can boost classification accuracy by up to 22% while reducing model parameters by 92.5%, from 14 million parameters from a baseline CNN model to 1.1 million parameters with the proposed MobileViTV2 model. Although this research focuses on rice leaf disease classification in precision agriculture, the techniques discussed are applicable to a wide range of image classification problems, particularly in resource-constrained environments like mobile devices and edge computing.

In this talk, I will cover:

1. Introduction to Image Classification in Precision Agriculture (5 mins): Why efficient models are crucial for real-time mobile deployment in agriculture.
2. Challenges with Traditional CNNs (7 mins): Limitations in computational resources and mobile deployment.
3. Introduction to MobileViTV2_050 (10 mins): How it combines CNNs and Vision Transformers for improved performance and efficiency.
4. Implementation Walkthrough (12 mins): A detailed explanation of how we trained and evaluated our enhanced models, with performance metrics.
5. Broader Applications and Next Steps (5 mins): How these techniques can be applied to other industries and research areas.

By the end of the session, participants will be equipped with the knowledge to implement transfer learning strategies using MobileViTV2 for mobile and edge-based image classification tasks, improving both accuracy and efficiency in their own projects.

The source code and research papers are available at https://drive.google.com/drive/folders/16ePM-kPq7BuZzoTQKvW_mQ-qZqfxYBLV?usp=sharing, and previous presentations can be found here in a form of CV: https://drive.google.com/file/d/1DBpHIWX8xenodrHJBi0fVQhmGkAmQm4L/view?usp=sharing.

Speakers experience:
As a 4th-year Computer Science student at Pamantasan ng Lungsod ng Maynila and a student tech leader for the AWS Cloud Club Haribon, I am passionate about Python and AI-driven innovations, particularly in Generative AI and computer vision. I have gained experience through engaging in webinars and seminar talks, including those hosted by DevCon Manila, the ISACA Manila Chapter, and AWS Cloud Club Haribon. Additionally, my team and I participated in the Generative AI hackathon, where we won first place at the AWS Innovation Cup 2024 by creating "Agap," an AI-based emergency triage system built with Streamlit.

Currently, I am working on my thesis, where I am leveraging the MobileViTV2 Vision Transformer to develop a mobile application for detecting rice leaf diseases. This project showcases how cutting-edge AI technologies can make a transformative, real-world impact, particularly in agriculture.

As a 2023 DOST JLSS Merit Scholar, I am deeply committed to advancing technology across various fields. I am eager to share my insights and experiences with the PyCon APAC 2025 community, especially in how tools like Generative AI, cloud computing, and machine learning can reshape industries ranging from healthcare to agriculture. I look forward to contributing to the growing field of Python development and engaging with like-minded individuals at PyCon APAC 2025.