RealWaste CNN Image Classification

EN3150 Assignment 03 - Pattern Recognition
A comprehensive deep learning implementation for multi-class waste image classification using Convolutional Neural Networks.

🎬 Live Demo

---

Dataset Overview

RealWaste Dataset from UCI Machine Learning Repository

• Total Images: 4,752 images (224×224 pixels)
• Classes: 9 waste categories
  • Cardboard, Glass, Metal, Organic, Paper, Plastic, Textile, Trash, Wood
• Data Split: 70% Training (3,326) | 15% Validation (712) | 15% Testing (712)
• Download: https://archive.ics.uci.edu/dataset/908/realwaste

---

Architecture & Models

1. Custom CNN (From Scratch)

Progressive architecture with hierarchical feature learning

Component	Configuration
Conv Blocks	4 blocks with filters: 32 → 64 → 128 → 256
Kernel Size	3×3 (Industry standard)
Pooling	2×2 MaxPooling after each block
Regularization	Batch Norm + Progressive Dropout (0.25→0.50)
Dense Layer	256 units with ReLU activation
Output	Softmax (9 classes)
Total Parameters	13.43M

Performance:

• Test Accuracy: 71.45% ✓
• Validation Accuracy: 73.30%
• Test Loss: 0.8587
• Training Accuracy: 79.83%

2. Transfer Learning Models

Model	Accuracy	Parameters	Type
MobileNetV2	76.56%	2.59M	Best Efficiency
Custom CNN	71.45%	13.43M	From Scratch
VGG16	68.47%	14.85M	Baseline

---

Key Experimental Findings

Learning Rate Analysis

Systematic testing with three learning rates over 15 epochs:

Learning Rate	Best Val Acc	Status
0.0001	34.52%	Too Conservative
0.001	64.06%	✓ Optimal
0.01	44.18%	Too Aggressive

Selected: 0.001 (balanced convergence speed and stability)

Optimizer Comparison

Impact of momentum in SGD-based optimization:

Optimizer	Test Accuracy	Observation
SGD (No Momentum)	11.79%	❌ Failed (random guessing)
SGD + Momentum	51.42%	+336% improvement
Adam	73.30%	✓ Best long-term

Key Insight: Momentum is essential for non-convex optimization; Adam provides superior adaptive per-parameter scaling.

Training Configuration

Optimizer:  Adam (lr=0.001, β₁=0.9, β₂=0.999)
Loss:       Sparse Categorical Cross-Entropy
Batch Size: 32
Epochs:     50 (Early stopping at epoch 42)
Callbacks:  ReduceLROnPlateau + EarlyStopping

---

Activation Functions

Layer	Function	Reason
Hidden Layers	ReLU	Avoids vanishing gradients, computationally efficient
Output Layer	Softmax	Multi-class probability distribution

---

Key Results Summary

Custom CNN Training

• Best Validation Accuracy: 73.30%
• Test Accuracy: 71.45%
• Generalization Gap: 6.53% (controlled overfitting)
• Training-Validation Balance: Effective regularization achieved

Performance Insights

✓ Progressive filter expansion captures hierarchical features efficiently
✓ Batch normalization stabilizes training and enables higher learning rates
✓ Progressive dropout (0.25→0.50) prevents overfitting in deeper layers
✓ Adam optimizer with LR=0.001 provides stable convergence

---

👥 Team ByteBrains

• S. Abisan (220013N)
• S. Changeethan (220084F)
• S. Pingalan (220478R)
• M. Thiruvarankan (220647K)

---

Citation

Single, S., Iranmanesh, S., & Raad, R. (2023). RealWaste [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5SS4G

---

License

This project is part of EN3150 Pattern Recognition course assignment.

Repository: RealWaste-CNN-Classfication