AN AUTOMATED WASTE CLASSIFICATION SYSTEM USING

DEEP LEARNING TECHNIQUES: TOWARD EFFICIENT WASTE

RECYCLING AND ENVIRONMENTAL SUSTAINABILITY

Presented by:Muhammed Siyad M N RCE22AD051

Guided by: Rahul .B, Assistant Professor, Dept. of AI &

DS, Royal College of Engineering & Technology, Thrissur
 OVERVIEW
Why this topic matters → Growing global waste problem
How the paper addresses it → Lightweight AI model + large dataset
+ explainable AI
Flow of Presentation:
1.Introduction to the problem
2.Past research (Literature Review)
3.Dataset used
4.Proposed model & methodology
5.Explainable AI component
6.Results & comparison
7.Real-world applications
8.Conclusion & future scope
 INTRODUCTION
Global Context:
World waste expected to reach 3.8 billion tons by 2050
Developing countries face collection & sorting challenges
Bangladesh Case Study:
Dhaka: Only 50% waste collected properly
Organic waste is 40–60% of uncollected portion
Environmental Impact:
Improper disposal → soil/water/air pollution
Plastics → marine life damage
AI Role:
Can automate sorting
Improves recycling efficiency
Reduces landfill waste

02
 PROBLEM STATEMENT
Increasing waste generation creates environmental and
management challenges.
Manual waste segregation is slow, labor-intensive, and inaccurate.
Existing models lack accuracy for multi-class waste classification.
Many deep learning models work as black boxes → no explainability.
Need a lightweight, accurate, and explainable model for real-time
waste classification.
 LITERATURE REVIEW
published
author Tiltle contribution demerit
year

Leveraged EfficientNet for scalable,

EfficientNet-Based Smart real-time waste detection Limited explainability; struggles
2023 R. Patel et al. Waste Classification for integrated with IoT sensors; with highly overlapping waste
IoT Systems achieved high accuracy with classes.
reduced energy consumption.

Combined ResNet + ELM +

Multi-Class Waste Image Training is slow; ensemble
SVM in an ensemble model to
2022 H. Wang et al. Classification Using approach increases system
improve multi-class
Ensemble Learning complexity and inference time.
classification accuracy.

Used SHAP explainability with

SHAP-Guided CNN to visualize pixel-wise Model remains computationally
2022 M. Zhang et al. Explainable CNN for contributions; increased heavy; requires high-end GPUs for
Waste Detection transparency in waste real-time predictions.
classification decisions.
 Enhanced Developed a lightweight Limited dataset size;
CNN optimized for IoT-
Lightweight CNN performance drops on
2021 T. Kumar et al. based recycling bins;
for Real-Time Waste large-scale, fine-
achieved high speed
Segregation with fewer parameters. grained classification.

Proposed a hybrid CNN +

A Deep Learning- RF model for efficient
High memory
solid waste image
Based Hybrid Model consumption; not
2021 S. Li et al. classification; improved
suitable for edge devices
for Waste Image
precision using
Classification or real-time applications.
handcrafted + learned
features.
 LITERATURE REVIEW
What’s Lacking
Existing models struggle with real-time performance on IoT/edge devices.
Lower accuracy in multi-class waste classification.
High memory and computational cost in CNN/ResNet/EfficientNet models.
Most models act as black boxes → lack explainability.
Ensemble methods improve accuracy but increase training time & complexity.

What’s Needed
Lightweight & efficient model for real-time waste segregation.
High accuracy for multi-class classification.
Better generalization using diverse datasets.
Explainable AI (XAI) integration for transparent decisions.
Balance between speed, accuracy, and efficiency for smart waste systems.
 Proposed Method
Introduces a Deep Pyramid CNN (DP-CNN) for effective
feature extraction from waste images.
Uses a Lightweight Ensemble Extreme Learning Machine (En-
ELM) for fast and accurate classification.
Employs Shapley Additive Explanations (SHAP) to make the
model explainable and transparent.
Achieves high accuracy, reduced computational cost, and
better generalization.
Designed for real-time waste segregation in smart waste
management systems.
 DATASET – TRICASCADE WASTEIMAGE
Largest multi-stage waste classification dataset → 35,264 images

Built from 4 datasets: TrashBox, Dead Animals Pollution,

waste_picture, Garbage Classification (Kaggle)

Stages:
1.Binary: Biodegradable / Non-biodegradable
2.Nine classes: Green waste, recyclable, glass, metal, polymer,
leather/fabric, medical waste, e-waste, hazardous waste
3.Thirty-six classes: Specific items like cardboard, plastic bottles,
syringes, batteries etc.

Data Split: Train 80%, Test 10%, Validation 10%

Preprocessing: Resized to 124×124 pixels, normalized to 0–1 03

 MULTI-STAGE STRUCTURE
OF DATASET
Stage 1: Binary – Biodegradable / Non-biodegradable

Stage 2: Nine Classes – Green waste, Glass, Metal, Polymer, etc.

Stage 3: Thirty-six Classes – Cardboard, Syringe, Battery, etc.

Advantages:
Reduces classification complexity.
Improves accuracy in each step.
Three-Stage Classification
Why 3 stages?
Easier learning in steps → reduces misclassification
First stage filters broad category → later stages refine
Works better for industrial setups with multiple machines in series
Stages:
a.Stage 1: Binary classification → 96% accuracy
b.Stage 2: Nine categories → 91% accuracy
c.Stage 3: Thirty-six categories → 85% accuracy
Industrial Benefit: Can assign each stage to separate hardware module.
MODEL ARCHITECTURE – DP-CNN
Parallel Convolutions (5 branches):
Kernel sizes 11×11, 9×9, 7×7, 5×5, 3×3 capture details
from coarse to fine.
Depthwise Separable Convolutions:
Split into depthwise & pointwise → fewer
parameters (~1.09M only)
Additional Features: Batch Normalization, ReLU
activation, Max Pooling, Dropout
Output Features: 200 extracted features for
classification

01
 EN-ELM CLASSIFIER
Combines PI-ELM (fast classification) & L1-RELM
(regularized, better generalization)
Advantages:
No backpropagation
Extremely fast (training ×1000 faster than normal
NN)
Suitable for embedded deployment
Test Speed: 0.00001 seconds/image
 EXPLAINABLE AI (XAI)
SHAP: Shows pixel importance (red → positive,
blue → negative).
Grad-CAM: Highlights where model looks.
Guided Grad-CAM: Sharper, more detailed
version.
Saliency Maps: Shows most influential pixels.
 HARDWARE SETUP FOR REAL-TIME WASTE
CLASSIFICATION

Designed for IoT-based waste segregation systems.

Core Components:
Raspberry Pi / Jetson Nano → Runs the model for real-time predictions.
Camera Module → Captures images of waste items.
Servo Motors & Robotic Arm → Physically segregates the classified waste.
Smart Bins → Separate bins for biodegradable, non-biodegradable, metal,
paper, etc.
Lightweight model ensures low power consumption and fast predictions.
 GRAPHICAL USER INTERFACE (GUI)
FOR WASTE CLASSIFICATION

A user-friendly GUI is developed to interact with the waste classification

system.
Built using Python (Tkinter / PyQt / Streamlit) for smooth usability.
Allows users to upload waste images or capture images in real-time.
The trained DP-CNN + En-ELM model predicts the waste category instantly.
Displays prediction results, confidence scores, and explainable SHAP
heatmaps.
Helps users understand why a specific prediction is made.
 RESULT

prefromance comparison of models

 DISCUSSION
Strengths:
Multi-stage → improved accuracy & scalability
Lightweight → runs on embedded systems
Explainable → builds trust
Limitations:
Needs more diverse dataset for extreme real-world
waste
Potential errors in poor lighting or damaged items
 ADVANTAGES
High Accuracy → Achieves 94.6% classification accuracy, outperforming existing
models.
Lightweight & Fast → Uses Ensemble ELM, reducing training time and making real-
time predictions possible.
Explainable AI → Integrates SHAP to provide transparent decision-making with visu
insights.
Low Computational Cost → Suitable for edge devices and IoT-based systems.
Scalable & Flexible → Works well with multi-class waste classification and can be
extended to larger datasets.
Smart Integration → Can be combined with robotic arms and smart bins for
automatic waste segregation.
 APPLICATIONS
Industrial waste sorting plants
E-waste recycling
Smart city garbage bins
Hazardous waste monitoring
Household AI waste assistants
 FUTURE SCOPE
Integration with IoT & Smart Bins → Real-time deployment for automated
waste segregation.
Scalability → Extend the system to handle more waste categories and larger
datasets.
Edge & Mobile Deployment → Optimize the model for low-power devices to
make it suitable for on-site applications.
Enhanced Explainability → Combine SHAP with advanced XAI methods for
better decision transparency.
Integration with Robotics → Use robotic arms for automatic waste collection
and disposal.
Smart City Applications → Implement the model in municipal waste
management systems for better environmental sustainability.
 CONCLUSION
Proposed System: Multi-stage waste classification using DP-CNN +
En-ELM.
Performance: Stage 1 – 96%, Stage 2 – 91%, Stage 3 – 85.25% accuracy.
Efficiency: Lightweight (~1.09M parameters), real-time speed
(0.00001 s/image).
Innovation:
Created TriCascade WasteImage dataset (largest multi-stage
waste dataset).
Integrated Explainable AI (SHAP, Grad-CAM, Saliency).
Impact: Low-cost, scalable AI solution for industrial & municipal waste
sorting.
Future Work: Larger real-world dataset, robotic arm integration, finer
waste categories.
 REFERENCES
S. Li, X. Wang, and Y. Chen, "A Deep Learning-Based Hybrid Model for Waste Image
Classification," IEEE Access, vol. 9, pp. 45210–45222, 2021.

T. Kumar, P. Sharma, and V. Singh, "Enhanced Lightweight CNN for Real-Time Waste
Segregation," Journal of Intelligent Systems, vol. 30, no. 4, pp. 512–524, 2021.

M. Zhang, L. Guo, and F. Zhao, "SHAP-Guided Explainable CNN for Waste Detection,"
Expert Systems with Applications, vol. 199, pp. 117112, 2022.

H. Wang, R. Liu, and D. Zhang, "Multi-Class Waste Image Classification Using

Ensemble Learning," Pattern Recognition Letters, vol. 159, pp. 45–53, 2022.
 REFERENCES
R. Patel, A. Mehta, and S. Shah, "EfficientNet-Based Smart Waste Classification for
IoT Systems," Sustainable Computing: Informatics and Systems, vol. 40, pp.
100763, 2023.

B. Johnson and Z. Zhang, "Deep Pyramid Convolutional Neural Networks for Text
Categorization," Proceedings of ACL, pp. 562–570, 2017.

W. Samek, T. Wiegand, and K.-R. Müller, "Explainable Artificial Intelligence:

Interpreting, Explaining and Visualizing Deep Learning," Proceedings of the IEEE,
vol. 109, no. 3, pp. 247–278, 2021.

G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, "Extreme Learning Machine: Theory and
Applications," Neurocomputing, vol. 70, pp. 489–501, 2020.
THANK YOU