Waste Classification Neural Network (WasteNet)

Automated 6-Category Recyclable Waste Classification with Vision Transformers & Transfer Learning

View Repository Contact Martin

Discipline Deep Learning & Computer Vision

Role Deep Learning Engineer

Timeline / Year 2025

Organization / Context University Deep Learning Project

Executive Overview

WasteNet is a comprehensive deep learning and computer vision system engineered to automate the sorting of recyclable materials into 6 distinct categories: cardboard, glass, metal, paper, plastic, and general trash. The project methodically benchmarks 5 progressive neural network architectures—from a custom convolutional baseline to state-of-the-art Vision Transformers (ViT)—culminating in a real-time Flask inference server and interactive web dashboard.

The Challenge & Solution

The Core Problem: Municipal recycling facilities and automated sorting plants face high contamination rates and manual sorting bottlenecks. Traditional computer vision techniques struggle with lighting variations, crumpled packaging, and reflections on transparent plastics.

The Engineering Solution: Designed an end-to-end deep learning pipeline incorporating sophisticated data augmentation and evaluated baseline CNN, deeper CNN with residual shortcuts, transfer learning (ResNet50 / MobileNetV2), and Vision Transformers (ViT-B/16). The best-performing model was packaged into an optimized Flask inference engine with live webcam feed classification.

System Architecture & Data Flow

Data Engineering

Collected and cleaned multi-source waste imagery; performed stratified train/validation/test splits.

Augmentation Pipeline

PyTorch torchvision transforms simulating real-world industrial conveyor lighting and deformations.

Progressive Benchmarking

Iterative training from baseline CNN (62% acc) to Transfer Learning (89% acc) to Vision Transformer (94.2% acc).

Inference Deployment

Local Flask microservice serving real-time predictions with top-3 confidence scores and recycling recommendations.

Key Engineering Highlights

Multi-Architecture Benchmarking

Detailed comparative analysis across 5 deep learning paradigms evaluating parameter efficiency vs accuracy.

Vision Transformer Integration

Implemented self-attention mechanisms capable of capturing global context across deformed waste items.

Real-Time Webcam Inference

Interactive web dashboard capable of processing 30 FPS video feeds on consumer GPUs.

Comprehensive Error Analysis

Generated confusion matrices and Grad-CAM activation maps highlighting model decision regions.

Technical Challenges & Resolutions

Engineering Challenge

Class Imbalance in Recyclables: Cardboard and paper had significantly more training samples than transparent glass.

Architectural Solution

Utilized weighted cross-entropy loss functions and minority-class synthetic augmentation to balance gradients.