Down-samples feature maps, reducing dimensions while preserving key features
Dense Layers
Fully connected neurons combine features for final classification
Softmax Output
Converts scores to probability distribution across waste categories
Convolutional Operations
The convolutional layer uses learnable filters (kernels) to scan the input image, generating output feature maps that detect spatial patterns.
Key Concepts
Local Connectivity: Each neuron connects only to a small region (receptive field)
Weight Sharing: Same filter weights used across all spatial positions
Translation Invariance: Detects features regardless of position in image
Multi-Channel Processing: Analyzes RGB channels simultaneously for color-dependent patterns
Benefits
Reduces learnable parameters compared to fully connected layers, improving computational efficiency and reducing overfitting risk.
Pooling Operations
Purpose
Pooling layers perform down-sampling to reduce spatial dimensions while maintaining essential features.
Max Pooling
Most commonly used function due to its ability to effectively map features within an image. Selects maximum value from each region.
Advantages
Reduces computational load
Provides translation invariance
Prevents overfitting
Maintains important features
Transfer Learning Approach
Base Models Tested
EfficientNetB0
EfficientNetV2 variants
ResNet50
MobileNetV2
All models pre-trained on ImageNet, enabling complex feature learning from extensive datasets.
Training Strategy
First half of base model layers frozen to preserve learned features. New classification head added for 6-class waste problem.
Adam optimizer with 1e-4 learning rate, categorical cross-entropy loss function, 15-30 epochs with early stopping.
Class weights computed to handle imbalanced dataset, ensuring fair representation of all waste categories.
Model Training Results
EfficientNetB0 achieved the highest classification accuracy of 91.97% on the test dataset, exceeding the 90% target. It offers the best trade-off between speed and accuracy for deployment.
Performance Metrics
91.97%
Test Accuracy
EfficientNetB0 model performance on held-out test set
0.91
F1-Score
Balanced measure of precision and recall across all classes
0.18
Validation Loss
Low loss indicates good model convergence and generalization
Precision and recall values exceeded 0.90 for main categories, confirming consistent performance. However, the model requires further training on Mongolian-specific waste data for reliable deployment.
Web Application Integration
QR Code Activation
When idle, Raspberry Pi displays QR code on screen. Scanning activates the system and prepares components for waste processing.
Direct camera activation available through web interface for seamless user experience.
Tracks individual contributions and classroom affiliations. Links users to their total recycling weight and classroom rankings.
2
Devices & Bins
Maintains device locations, serial numbers, and bin configurations. Monitors device health and last connection times.
3
Reports
Records every disposal event with classification, weight, image URL, and user ID. Enables analytics and leaderboard generation.
4
Misclassifications
Stores human corrections of AI errors. Creates feedback loop for continuous model improvement.
Mechanical Design Requirements
Structural Integrity
Robust construction to withstand outdoor conditions, resisting corrosion, impact, and fatigue for long-term operation.
Precise Actuation
Accurate sorting mechanisms with controlled movement for reliable separation of different waste categories.
User Safety & Accessibility
Features to prevent injury during interaction, including clear interfaces and easily accessible emergency stop mechanisms.
Weather Resistance
Enclosures and components rated for extreme temperatures, precipitation, and UV exposure for reliable outdoor deployment.
Maintenance Accessibility
Modular design allowing easy access to internal components for quick repairs, cleaning, and preventative maintenance.
Size Constraints
Optimized dimensions for efficient space utilization in various urban and rural deployment scenarios.
Mechanical Configuration
The input bin and integrated weighing system underwent extensive design and testing through multiple iterations.
1 moving stepper design
2 servo design
1 stepper with multi servo design
1 servo 1 stepper design
1 servo with 2 stepper design
This iterative approach allowed us to refine the mechanical configuration for optimal performance, durability, and user interaction, resulting in the current design of the prototype.
Figure shows a 3D design drawn on cad.onshape.com that was not completed
Other smaller parts were 3D printed uniquely custom to the project.
Key Achievements
Classification Success
CNN model achieved 91.97% accuracy, exceeding the 90% target requirement
Dataset Creation
Established first public Mongolian waste image database with 400+ labeled images
Modular Architecture
Developed scalable system supporting multiple deployment scenarios and configurations
Web Platform
Created functional reward system with real-time tracking and classroom leaderboards
ESP-32 code functions
Main Functional Flow:
01
Lift up
02
Measure weight
03
Lower lift
04
Move box
05
Open latch
06
Return to positions
System Features & Control Interface
Weighing Functionality
Automated lift sensor positioning for precise measurements
Real-time weight sampling with HX711 and data averaging
Tare functionality for empty box weight calibration
Motor Control
Non-blocking stepper motor movements for simultaneous operations
Horizontal box positioning and vertical lift control
Precise positioning with step-based movement tracking
Interactive serial command interface for manual control
Status reporting for system state monitoring
Current Limitations with future improvements
Partial Integration
System modules functional individually but not yet assembled into unified physical device due to late hardware arrival.
Dataset Size
Mongolian waste dataset still limited (400 images). Model requires retraining with more local data for reliable deployment.
Field Testing
No long-term operational validation yet. System performance under continuous real-world use remains to be evaluated.
Mechanical Design
Enclosure manufacturing and full mechanical assembly pending. Weather protection for outdoor deployment not yet implemented.
Electrical configuration
The final product will be using a custom printed PCB with IR sensors and Limit switches, and possibly other sensors such as NIR.
Connectivity
The model will be recieving A7670E SIM module for LTE connection incase of no WiFi availibility.
Future Work plan until June
Dataset Expansion
Continue collecting Mongolian waste images through public web interface. Target 2000+ images for robust model retraining.
Physical Assembly
Complete hardware integration into cohesive prototype. Manufacture weatherproof enclosure for outdoor deployment.
GMIT Pilot
Deploy system at GMIT to observe user behavior, system reliability, and long-term data consistency.
Optimization
Enhance communication stability, power efficiency, and cloud scalability for broader deployment across schools.
Thank You
This project demonstrates how engineering, AI, and human-centered design can converge to tackle Mongolia's urban waste challenges.
"The Automated Trash Sorting System stands not only as a technical prototype but also as a stepping stone toward building a culture of responsible waste handling in Mongolia."