Overview

This project is a joint research collaboration between Ritsumeikan University (Japan) and Universitas Dinamika Surabaya (Indonesia).
We developed an automated tomato quality assessment system using computer vision and machine learning to improve the efficiency, consistency, and scalability of traditional grading processes.

Imgur

Collaboration Structure

This research was conducted through an international collaboration with clearly defined roles:

  • Ritsumeikan University (Japan)

    • Led the system design and development
    • Implemented the machine learning models and pipeline
    • Conducted experiments and performance evaluation

  • Universitas Dinamika Surabaya (Indonesia)

    • Led data collection and dataset preparation
    • Collected real-world tomato images from local environments
    • Performed manual labeling and data annotation

This division of responsibilities enabled us to combine strong technical implementation with real-world data acquisition, improving the practical relevance of the system.

Motivation & Background

Tomato grading is typically performed manually, which introduces:

  • Human bias and inconsistency
  • High labor costs
  • Limited scalability

This project aims to solve these issues by building an automated, image-based grading system.

System Architecture

1. Image Acquisition

  • Individual input (uploaded images)
  • Industrial input (video frames from conveyor systems)

2. Preprocessing

  • Segmentation to isolate the tomato
  • Auto-cropping to remove background noise
  • Normalization for consistent input conditions

Feature Extraction

  • Color Features: RGB distribution, ripeness indicators
  • Shape Features: Size, circularity, symmetry
  • Texture Features: Surface smoothness, defect detection

Machine Learning Models

Freshness Classification (CNN)

  • VGG16-based architecture
  • Transfer learning and fine-tuning
  • Captures high-level visual features related to ripeness

Physical Quality Assessment (SVM)

  • Uses handcrafted features
  • Evaluates uniformity, size, and defects

Output

  • Freshness Grade: A–D
  • Physical Quality Score: 1–4

Dataset & Experiments

  • ~10,000+ training images (open-source datasets)
  • ~300 real-world images collected and labeled by Universitas Dinamika
  • Evaluation using:
    • F1 Score
    • 5-fold cross-validation

Results

  • Successfully built an end-to-end grading system
  • CNN effectively captured freshness-related features
  • SVM provided interpretable physical quality evaluation
  • Hybrid approach improved overall robustness

Industrial Extension

  • Video-to-frame processing for batch grading
  • Designed for real-time use in conveyor-based systems
  • Demonstrated scalability for agricultural and industrial applications

Limitations

  • Dataset diversity remains limited
  • Freshness classification partially relies on mapping
  • Limited number of grading classes

Future Work

  • Expand dataset with more environmental variations
  • Train models from scratch for improved accuracy
  • Develop a user-friendly interface
  • Optimize for real-time deployment

Tech Stack

  • Python, OpenCV, NumPy, Pandas
  • TensorFlow, PyTorch, Scikit-learn
  • Matplotlib, Seaborn

Key Contributions

  • Developed a hybrid ML system (CNN + SVM)
  • Designed a full end-to-end pipeline
  • Contributed to an international collaborative research project
  • Bridged technical development and real-world data collection

Takeaways

Through this project, I gained experience in:

  • International research collaboration
  • End-to-end AI system development
  • Computer vision for real-world applications
  • Data-driven problem solving

Impact

This project demonstrates how AI can be applied to agriculture to:

  • Improve efficiency and consistency
  • Reduce human bias
  • Enable scalable quality assessment systems

It reflects my interest in applying AI and optimization techniques to real-world problems in a global context.