AI Waste Sorter for Recycling Accuracy

ISEF Category: Environmental Engineering

Subcategory: Recycling and Waste Management  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

One wrong item can contaminate a whole recycling bin. That means one greasy box, one plastic bag, or one metal cap can send good material to the landfill. You can test whether a phone camera and an AI model can do better than a single-stream bin.

What Is It?

This project asks a simple question, can a computer vision model tell different waste items apart well enough to improve recycling sort quality? Computer vision means teaching a computer to find patterns in images. Transfer learning means you start with a model that already knows basic image features, then retrain it on your own waste photos.

Think of it like training a new helper who already knows what edges, shapes, and textures look like. You are not teaching the model from zero. You are teaching it how to recognize your local waste stream, such as cans, cardboard, glass, and common contaminants. Your local dataset matters because waste items vary by region, brand, lighting, and condition.

Why This Is a Good Topic

This is a strong science fair topic because you can measure it with real data. You can test classification accuracy, contamination-rate reduction, and confusion between similar items. The project connects to a real waste problem that cities, schools, and recycling centers face every day. You can also make it more original by using photos from your own community instead of a generic online dataset.

Research Questions

• How does a transfer-learning ResNet model compare with a simple baseline classifier for sorting local waste images?
• What is the effect of adding more locally photographed samples on classification accuracy?
• Does image background clutter change the model's ability to identify recyclable items?
• To what extent does lighting variation reduce the model's performance on phone photos?
• Which waste categories produce the most false positives and false negatives?
• How does model confidence relate to actual contamination-rate improvement in a simulated recycling bin?

Basic Materials

• Smartphone camera with a consistent photo setup
• Laptop or desktop computer with internet access
• Small set of clean sample waste items from your local stream
• Plain background board or photo box
• Labels or sticky notes for category tagging
• Spreadsheet software for organizing images and results
• Digital scale for tracking sample counts by category.

Advanced Materials

• Access to a school or university computer with a GPU
• Python environment with TensorFlow or PyTorch
• Pretrained ResNet model weights
• Image labeling tool such as LabelImg or CVAT
• External storage for a larger image dataset
• Controlled lighting setup for repeatable image capture
• Confusion matrix and statistical analysis software.

Software & Tools

• Python: Runs the training, testing, and evaluation code for your image model.
• TensorFlow: Supports transfer learning with pretrained vision models like ResNet.
• PyTorch: Offers flexible model training and easier custom testing workflows.
• Google Colab: Lets you train models in a browser if your own computer is slow.
• ImageJ: Helps inspect image quality and compare lighting or color differences across photos.

Experiment Steps

1. Define the waste categories you will sort, and decide which items count as contamination.
2. Build a photo protocol so each item gets captured in a repeatable way.
3. Split your dataset into training, validation, and test groups before you train anything.
4. Train a baseline model first, then compare it with a transfer-learning ResNet model.
5. Measure performance with a confusion matrix, not just overall accuracy.
6. Translate model outputs into a recycling metric, such as contamination-rate reduction in a simulated stream.

Common Pitfalls

• Mixing categories that look similar, such as clean paper and coated paper, which makes the model learn the wrong labels.
• Taking photos under changing light, which shifts color and hurts model consistency.
• Letting one item appear many more times than the others, which biases the classifier toward the largest class.
• Testing on photos that look too much like the training set, which inflates accuracy and hides weak real-world performance.
• Using vague contamination labels, which makes it hard to connect the model's predictions to recycling outcomes.

What Makes This Competitive

A competitive version of this project goes beyond a simple image classifier. You can test whether local data, lighting control, and category design change real recycling performance. You can also compare multiple models, report class-level errors, and connect predictions to contamination-rate improvement instead of only accuracy. That makes your project feel like an engineering study, not just a coding demo.

Project Variations

• Train the model on school cafeteria waste instead of household waste to study a different local stream.
• Compare RGB photos with grayscale or cropped object-only images to see which inputs improve sorting.
• Add a human-sorter baseline and measure whether the model beats student volunteers on the same test set.

Learn More

• MIT OpenCourseWare: Search for computer vision and machine learning course materials that explain image classification and transfer learning.
• NIH PubMed: Search review articles on waste sorting, recycling contamination, and computer vision in environmental monitoring.
• NASA Earthdata: Explore image analysis and remote sensing examples that build skills in pattern recognition and data handling.
• Pattern Recognition and Machine Learning by Christopher Bishop: A widely used textbook that covers classification, training, and evaluation concepts.
• IEEE Xplore: Search for peer-reviewed papers on waste classification, recycling sorting systems, and transfer learning in computer vision.