https://github.com/moatifbutt/waste-classification

https://github.com/moatifbutt/waste-classification

Last synced: over 1 year ago
JSON representation

• Host: GitHub
• URL: https://github.com/moatifbutt/waste-classification
• Owner: moatifbutt
• Created: 2021-02-08T13:23:17.000Z (over 5 years ago)
• Default Branch: main
• Last Pushed: 2021-02-08T17:17:52.000Z (over 5 years ago)
• Last Synced: 2025-02-09T09:43:58.390Z (over 1 year ago)
• Language: Jupyter Notebook
• Size: 1.65 MB
• Stars: 0
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# Waste Classification Using Convolutional Neural Network

## Pre-Requisites

1. Anaconda (https://www.anaconda.com/products/individual)

2. Pytorch (https://pytorch.org/get-started/locally/)

3. Waste Classification Binary Class Dataset (https://www.kaggle.com/techsash/waste-classification-data)

4. Thung and Wang's Dataset (https://github.com/garythung/trashnet/blob/master/data/dataset-resized.zip)

## Data Augmentation

To add more diversity in dataset, we have performed three types of data augmentations including,

- Gaussian Blur

- Horizontal Flip

- Contrast

You can use the **Augmentation.ipynb** in **Train_Notebooks** to perform augmentation on the provided dataset.

## Training

To this end, three state-of-the-art CNNs i.e. **GoogleNet** , **Resnet-50**, **VGG**, and **ResNext** are employed to perform classification. Based on the best performed, we selected **ResNext**, and added a new classification block has been added in the basis of network. It is important to mention here that, pre-trained CNNs are downloaded, and trained through transfer learning to achieve effective results. All the networks are available in above-folder named **Trained Notebooks**. All the trained models are available in above-folder **models**.

## Results

The experiments are performed on both i.e. the augmented and non augmented datasets to analyse the difference through adaptive and static learning rate. The experiments showed that the modified ResNext achieved better accuracy on augmented dataset with adaptive learning rate.

#### GoogleNet

- With Adaptive Learning Rate on non-Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/google-non.PNG?raw=true)

- With Adaptive Learning Rate on Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/google-aug.PNG?raw=true)

#### Resnet

- With Adaptive Learning Rate on non-Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/res-non.PNG?raw=true)

- With Adaptive Learning Rate on Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/res-aug.PNG?raw=true)

#### VGG

- With Adaptive Learning Rate on non-Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/vgg-non.PNG?raw=true)

- With Adaptive Learning Rate on Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/vgg-aug.PNG?raw=true)

#### ResNext

- With Adaptive Learning Rate on non-Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/resnxt-non.PNG?raw=true)

- With Adaptive Learning Rate on Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/resnxt-aug.PNG?raw=true)

#### ResNext-Plus

- With Adaptive Learning Rate on non-Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/resnxt-non.PNG?raw=true)

- With Adaptive Learning Rate on Augmented Dataset

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/respls-aug.PNG?raw=true)

#### ResNext-Plus on Sencond Dataset (Thung and Yang’s)

![alt text](https://github.com/aatiibutt/waste-classification/blob/main/graph/resnxtpls-seconddata.PNG?raw=true)