projectCicleR

In collaboration with Dhruv Hariharan

Introduction

This github repository contains all the code for our projectCicleR. This project focuses on a solution to reducing waste from the landfills by segregating organic and recyclable garbage with the help of Computer vision and IoT.

Webapp

We also built a webapp for easy access to the deep mearning model. The web app is accessible at http://projectcicler.pythonanywhere.com/

This repository also contains the code for the same which was built using django. To set up the webapp in your local system, clone this repository and install the requirements in the requirements.txt.

Some of the major requirements for the webapp are

1. Django
2. Tensorflow
3. numpy
4. basic_classifier.h5 (this is available in models/basic_classifier.h5)

Please note that the DJANO_SECRET_KEY has been hidden and will have to be generated to run correctly in the local system.

To create a new DJANGO_SECRET_KEY use the following command

from django.core.management.utils import get_random_secret_key
get_random_secret_key()

The generated key can then be stored inside core/.env (Please create this file in the specified directory)

CNN Model explained

For segregating the waste, we built a Convolutional Neural Network with three convolutional layers and three layers in the feed forward network. Even though this is a simple model, we were able to achieve a raw accuracy of around 88%.

The code for building the CNN model is inside a Jupyter Notebook. The notebook is located in scripts/training_with_tf2.ipynb

Dataset

For the dataset, we scraped images from the internet of organic and recyclable waste. The dataset contains 22564 training set and 2513 validation set. The dataset can be accessed from this link

Future Work

1. Dataset: The dataset can be improved by adding more images pertaining to organic and recyclable materials and cleaning the current dataset. This will also help increase the accuracy of the model.

2. Webapp: We intend to fine-tune the website by adding more features, improving UI, making the app more user friendly and more. Make it more responsive.

3. Image Classifier: The current model is a basic model built on limited resources and time. We fully intend to increase complexity, include reinforced learning and improve accuracy.