EcoSort: AI-Powered Waste Classification

EcoSort is a web-based application that uses computer vision to classify waste materials and provide recycling guidance. I grew up in a household that took recycling seriously, and even with that background, I learned how confusing, and sometimes daunting, it can be to sort things correctly. Will I make the right call? Could I be dooming the planet? EcoSort is one small step toward making that decision easier: snap a photo and get a clear bin recommendation backed by AI.

Model Architecture

Core Model: MobileNetV3-Small

The project uses MobileNetV3-Small as the backbone architecture, chosen for its optimal balance of accuracy and efficiency for mobile/web deployment:

• Architecture: MobileNetV3-Small with ImageNet pre-trained weights
• Input Size: 224×224 pixels (standard ImageNet resolution)
• Output: 6-class classification (glass, paper, cardboard, plastic, metal, trash)
• Optimization: ONNX format for cross-platform deployment
• Inference: Client-side using ONNX Runtime Web with WASM

Classification Hierarchy

The model implements a two-level classification system:

Material Classification (6 classes)
├── Glass → Recycle
├── Paper → Recycle  
├── Cardboard → Recycle
├── Plastic → Recycle
├── Metal → Recycle
└── Trash → Landfill

Training Configuration

• Data Augmentation: Color jittering, Gaussian blur, random transformations
• Normalization: ImageNet mean/std values ([0.485, 0.456, 0.406] / [0.229, 0.224, 0.225])
• Loss Function: Cross-entropy loss
• Optimizer: Adam with learning rate scheduling
• Validation Split: 10% stratified split

Data Construction

Dataset Structure

The training dataset consists of 2,527 images organized into 6 material categories:

data/raw/
├── glass/    
├── paper/      
├── cardboard/ 
├── plastic/   
├── metal/     
└── trash/     

Data Processing Pipeline

1. Manifest Creation: data/create_manifest.py processes the raw dataset

   • Parses zero-indexed-files.txt with filename-label pairs
   • Maps material IDs to bin categories
   • Generates CSV manifest with paths and labels

2. Stratified Splitting: Ensures balanced representation across classes

   • Training: 90% of data
   • Validation: 10% of data
   • Test: Optional split for final evaluation

3. Preprocessing: Standard ImageNet normalization and augmentation

   • Resize to 224×224
   • Color jittering and blur for training
   • Tensor normalization

Web Application

EcoSort is hosted on Cloudflare at https://eco-sort.pages.dev/.

Key Features

• Dual Input Methods: File upload + camera capture
• Real-time Processing: Instant classification results
• Responsive Design: Works on desktop and mobile

Basic Usage

1. Open the app in your browser
2. Upload a photo or take a picture using your camera
3. View results with confidence percentage and bin recommendation
4. Follow recycling tips provided for each material type

Example Classifications

High Confidence Example

Scenario: Clear plastic water bottle

• Result: RECYCLE · plastic (94.9%)

Low Confidence Example

Scenario: Unrelated Picture

• Result: UNSURE (52.3%)

Flyer & QR Code

With the goal of promoting sustainability and proper practices the following flyer (EcoSort_Flyer.png), advertises the hosted web app via a QR code. With the goal of being posted where people must make recycling descisions, at stores, schools, dumps, etc, EcoSort hopefully, in due time, will make a small contribution to saving the world.

Next Steps

• Refine and optimize the model.
  • Gather more images to train on
  • Account for certain types of plastics and cardboard that can not be recycled (ie pizza boxes (due to grease))
• Expand the UI
  • Allow for user to critic model results / report incorrect assumptions which then save the photo
  • Allow for users to submit photos of items, with proper labeling