NeuroBin: AI-Powered Smart Waste Segregation System

NeuroBin is a smart waste classification and segregation system that integrates deep learning and embedded automation to solve the challenge of real-time waste disposal. The system classifies waste as Biodegradable or Non-Biodegradable and triggers a physical mechanism to open the appropriate bin.

Project Overview

NeuroBin is an AI-powered waste segregation system designed to automate the classification of waste at the point of disposal. It uses a Convolutional Neural Network (CNN) to identify whether an item is Biodegradable or Non-Biodegradable based on its visual characteristics. This classification then triggers a corresponding physical response — the automatic opening of the appropriate dustbin lid.

The primary objective is to reduce improper waste disposal and promote recycling and sustainability through intelligent decision-making at the bin level. The solution is scalable and can be integrated into smart homes, schools, institutions, and city-wide smart bins.

This project demonstrates how the integration of machine learning with real-world embedded systems can create practical and impactful environmental solutions. The complete solution involves:

• A custom-trained CNN using more than 250,000 waste images
• Fine-tuning using structured and curated datasets
• Real-time inference via camera feed and image classification
• Integration with servo-controlled physical bins

Why NeuroBin?

• Real-time AI-powered waste classification
• Deployed on edge device (Raspberry Pi)
• Tested on real-world waste scenarios
• High accuracy with robust generalization

Circuit Diagram

Real-World Demonstration

• Apple → Biodegradable → Biodegradable(B) bin opens
• Cardboard → Biodegradable → Biodegradable(B) bin opens
• Amul Lassi Tetra Pack → Non-Biodegradable(N) → Red bin opens

Video and image proof of the live system operation are provided in the assets/ folder.

• Video Demo: assets/Demo_video.mp4
• Smart Bin Image: assets/neurobin.jpg

Datasets Used

1. rayhanzamzamy/non-and-biodegradable-waste-dataset
   Used for initial model training (~256,000 images).

2. sujaykapadnis/cnn-waste-classification-image-dataset
   Used for fine-tuning to increase model performance.

3. techsash/waste-classification-data
   Used as an independent evaluation/test dataset.

Model Architecture

The classification model is a custom-built Convolutional Neural Network (CNN) comprising:

• Three convolutional layers with max pooling
• A flattening layer followed by a dense hidden layer
• Dropout regularization for generalization
• Output layer with softmax activation for binary classification

The final model is exported in .keras and .tflite formats, making it compatible with both software testing and real-time embedded deployment.

Files Included

How to Run

1. Clone this repository:

git clone https://github.com/utkarshgupta-ai/NeuroBin-Smart-Waste-Segregator.git
cd NeuroBin-Smart-Waste-Segregator

2. Install the required packages:

pip install -r requirements.txt

3. Run training or demo scripts as needed:

python model_pretrain.py         # Train from scratch
python model_training.py         # Fine-tune pretrained model
python model_evaluation.py       # Evaluate model performance

For real-time testing:

• Open model_inference_demo.py in Google Colab
• Upload any waste image
• View predictions and classification confidence

Model Testing & Results

The NeuroBin model was tested on real-world waste items including food, packaging, and mixed materials.

Image	Prediction	Confidence
	Biodegradable	100%
	Biodegradable	99.98%
	Non-Biodegradable	68.78%
	Biodegradable	100%
	Biodegradable	73.21%
	Non-Biodegradable	87.97%
	Biodegradable	100%
	Non-Biodegradable	87.97%
	Biodegradable	100%
	Non-Biodegradable	99.98%
	Non-Biodegradable	100%
	Non-Biodegradable	99.96%

Key Observations

• Achieves near-perfect accuracy on organic waste items
• Strong classification performance on packaged materials
• Robust across real-world conditions and lighting variations

Model Performance

• Training Accuracy: 94.28%
• Validation Accuracy: 93.56%

Real-World Testing

The model was tested on diverse real-life waste scenarios including food items, plastic packaging, and mixed materials under varying lighting conditions.

The model demonstrates strong generalization across unseen real-world waste items.

Deployment

The final model (.keras or .tflite) can be embedded into a Raspberry Pi, Jetson Nano, or ESP32-CAM-based system using OpenCV and GPIO controls to activate bins via servo motors or relays.

Project Demo

Evaluation and Metrics

• Validation Accuracy: ~94%
• Loss Function: Sparse Categorical Crossentropy
• Optimizer: Adam
• Evaluation: Performed on a separate unseen dataset with high confidence

Confusion matrix and accuracy reports are generated via model_evaluation.py.

Author

Utkarsh Gupta
Domain expertise: Machine Learning, Embedded Systems, Computer Vision

License

This project is licensed under the MIT License. Feel free to use, modify, and distribute with attribution.