CoolFace
Apppublic

neecat/trash-classification-pytorch

sourceHugging Faceupdated 1y agoView on Hugging Face
2likes
App README

Trash Classifier with Explainability

A deep learning project to classify recyclable waste items from images, enhanced with visual explainability tools. This pipeline includes model training, evaluation, and a real-time interactive app to demonstrate classification and model interpretation.

Purpose and Why This Project Matters

Recycling is a crucial step toward sustainability, but it remains a manual and error-prone process. Automating trash classification helps:

  • —Improve waste sorting accuracy
  • —Reduce contamination in recycling streams
  • —Enable smart waste management systems

This project builds a lightweight yet powerful image classifier that recognizes 6 categories of recyclable and non-recyclable waste, enhanced with saliency-based interpretability to show why the model predicts what it does.

Architecture Overview

  • —Backbone Model: MobileNetV2 (pretrained on ImageNet)
  • —Classifier Head: Fully connected layer with 6 output classes
  • —Explainability: Captum's Saliency, SmoothGrad, Guided Backpropagation
  • —Deployment: Streamlit app for real-time prediction and visualization

Dataset Used

  • —Name: TrashNet
  • —Classes: cardboard, glass, metal, paper, plastic, trash
  • —Image Size: 512x384 pixels
  • —Class Distribution:
  • —Paper: 23.51%
  • —Glass: 19.83%
  • —Plastic: 19.07%
  • —Metal: 16.22%
  • —Cardboard: 15.95%
  • —Trash: 5.42%
  • —Handling: Dataset split into train, val, test folders with stratification

Techniques and Tools Used

  • —Model Training:
  • —Transfer learning with frozen base layers
  • —Adam optimizer, CrossEntropyLoss
  • —TensorBoard for live metric tracking
  • —Preprocessing:
  • —Resize & normalize images to 224x224
  • —Stratified train-val-test split
  • —Explainability:
  • —Captum's:
  • —Saliency: Base gradients
  • —SmoothGrad: Smoothed attribution
  • —GuidedBackprop: Edge-focused explanations
  • —Visualization: Matplotlib side-by-side comparisons
  • —Deployment:
  • —Streamlit app with file uploader and toggle for saliency types

Training Procedure

  • —Epochs: 10
  • —Optimizer: Adam (lr = 1e-4)
  • —Loss Function: CrossEntropy
  • —Early Stopping: Not needed (no overfitting)
  • —Logging: TensorBoard for accuracy/loss

Training monitored using:

  • —Accuracy/train and Accuracy/val
  • —Loss/train and Loss/val

Training Accuracy/Validation Graph:

<img src="evalpics/accloss.png" width="60%"/>

Result Evaluation

MetricValue
Train Accuracy98.1%
Val Accuracy91.1%
Train Loss0.071
Val Loss0.312

Evaluation Tools:

  • —TensorBoard visualization
  • —Saliency-based interpretation (SmoothGrad + Guided)
  • —Streamlit demo UI

Visual Placeholder:

<table> <tr> <td><img src="evalpics/streamlittestpic.png" width="100%"/></td> <td><img src="evalpics/saliencymap.png" width="100%"/></td> </tr> </table> ---

Overall Project Impact

  • —Demonstrates real-world machine learning application
  • —High performance on imbalanced, noisy data
  • —Includes interpretability and explainability tools
  • —Easy-to-use UI for non-technical users

This project shows a complete ML pipeline — from raw image data to deployed app with transparency in decision-making.

Final Result

A robust trash classifier trained on real data, achieving 91.1% accuracy on validation and equipped with visual explainability, ready to be deployed in sustainability-focused applications, smart bins, or recycling plants.

[Live Demo Link Coming Soon]

[Try Locally]:

bash
streamlit run web_app/app.py