A data science senior thesis project by Omar Moustafa, Malak Elsayed, & Nour Kahky
Department of Mathematics and Actuarial Science, The American University in Cairo (AUC)
Supervised by Dr. Noha Youssef · In collaboration with Nestlé and Dawar
This is an end-to-end deep learning pipeline that automatically analyzes incoming waste from a single image and classifies it across four critical dimensions:
| Task | Model Architecture | Validation Accuracy |
|---|---|---|
| Liquid Detection | MobileNetV2 | 94.7% |
| Single vs. Multi-Component Detection | MobileNetV2 | 94.5% |
| Recyclable vs. Non-Recyclable Classification | DenseNet121 | 96.8% |
| Waste Material Classification | Xception | 94.3% |
The four independently trained CNN models are integrated into a sequential decision pipeline that mirrors real-world bin behavior. On an external real-world validation dataset provided by Dawar, the unified pipeline achieved 92.4% accuracy.
This research was funded by Nestlé and Dawar, and culminated in a physical smart bin prototype built by AUC's Eltoukhy Learning Factory and MakersGate, integrating the deep learning pipeline with sensors, cameras, and embedded hardware for real-world deployment.
The pipeline processes each input image through four sequential stages, as illustrated by the following figure:
Figure 1. Sequential multi-stage SmartSort waste classification pipeline.
Decision Codes:
| Code | Meaning | Action |
|---|---|---|
| 1 | Liquid detected | Prompt user to empty liquid |
| 2 | Multi-component item | Prompt user to separate components |
| 3 | Recyclable — Plastic | Dispense to plastic compartment |
| 4 | Recyclable — Paper | Dispense to paper compartment |
| 5 | Recyclable — Organic | Dispense to organic compartment |
| 6 | Recyclable — Other | Dispense to general recycling compartment |
| 7 | Non-recyclable | Dispense to general waste compartment |
The repository is organized into the following main directories and files:
smartsort/
│
├── images/
│ ├── system_architecture.png # Pipeline architecture diagram
│ └── prototype_design.png # Physical smart bin prototype
│
├── models/ # ⚠️ Weights not included — see note below
│ ├── final_adjusted_liquid_advance.h5 # Liquid detection model (MobileNetV2)
│ ├── component_model_fixed_v3.h5 # Component detection model (MobileNetV2)
│ ├── DenseNet121_binary_v2.keras # Recyclability classifier (DenseNet121)
│ └── xception_4classes.keras # Material classifier (Xception)
│
├── notebooks/
│ └── smart_waste_bin_pipeline_may22.ipynb # Unified inference pipeline (Google Colab)
│
├── requirements.txt # Full dependency list
└── README.md
⚠️ Model weights are not included in this repository due to file size constraints. To run the pipeline, you will need the following four files:
final_adjusted_liquid_advance.h5component_model_fixed_v3.h5DenseNet121_binary_v2.kerasxception_4classes.kerasContact the authors for access.
- Python 3.9+
- TensorFlow 2.19.0
- Google Colab (recommended) or a local GPU environment
pip install tensorflow==2.19.0 opencv-python-headless numpy pillowOr install all dependencies from the full requirements file:
pip install -r requirements.txtThe main pipeline is contained in notebooks/smart_waste_bin_pipeline_may22.ipynb. It is designed to run in Google Colab.
- Upload the notebook to Google Colab.
- Upload all four model files to your Colab session (or mount Google Drive and update the model paths).
- Run all cells to load the models.
- When prompted, upload a waste image — the pipeline will run all four classification stages and return a decision code and action message.
# Model paths — update these if needed
LIQUID_MODEL_PATH = 'final_adjusted_liquid_advance.h5'
COMPONENT_MODEL_PATH = 'component_model_fixed_v3.h5'
BINARY_MODEL_PATH = 'DenseNet121_binary_v2.keras'
CATEGORY_MODEL_PATH = 'xception_4classes.keras'============================================================
PROCESSING: cola_bottle.jpeg
============================================================
🔍 Stage 1 · Liquid Detection
raw=0.8231 → ✅ No liquid (82.3%)
🔍 Stage 2 · Component Detection
raw=0.9104 → ✅ Single component (91.0%)
🔍 Stage 3 · Waste Classification
♻️ RECYCLABLE — PLASTIC (97.2%)
============================================================
🎯 DECISION CODE : 3
📢 ACTION : Dispensing to Plastic recyclable compartment
============================================================
- Architecture: MobileNetV2 (transfer learning)
- Task: Binary classification — liquid vs. non-liquid content
- Input: 224 × 224 RGB image
- Threshold: 0.5 (raw score < 0.5 → liquid detected)
- Architecture: MobileNetV2 (transfer learning)
- Task: Binary classification — single vs. multi-component item (e.g., paper cup with plastic lid)
- Input: 224 × 224 RGB image
- Threshold: 0.5 (raw score < 0.5 → multi-component detected)
- Architecture: DenseNet121 (transfer learning)
- Task: Binary classification — recyclable vs. non-recyclable
- Input: 224 × 224 RGB image, DenseNet preprocessing
- Architecture: Xception (transfer learning)
- Task: Multi-class classification — Paper, Plastic, Organic, or Other
- Input: 299 × 299 RGB image, Xception preprocessing
- Only runs if the recyclability model returns "recyclable"
No large-scale public dataset for multi-task waste classification existed at the time of this project. To address this, we utilized a combination of publicly available datasets, self-constructed datasets, and industry-provided external validation data. Each dataset was selected or created for a specific classification task within the pipeline, ensuring that the models were trained on data that reflects the complexities of real-world waste sorting.
Waste Material Classification Dataset
Publicly available on Kaggle, this dataset contains 11,522 labeled RGB images spanning nine waste categories: E-waste, Automobile Waste, Battery Waste, Glass Waste, Light Bulbs, Metal Waste, Organic Waste, Paper Waste, and Plastic Waste. For this project, the nine categories were remapped into four classes—Paper, Plastic, Organic, and Other—to balance class frequencies and improve model generalization. The "Other" category consolidates less frequently encountered materials such as E-waste, automobile waste, battery waste, glass waste, and light bulbs. All images were resized to 224×224 pixels and stored in .jpg format.
Recyclable vs. Non-Recyclable Dataset
Publicly available on GitHub, this dataset comprises 34,459 RGB images labeled as either Recyclable or Non-Recyclable. The images are organized into class-specific directories, enabling efficient data loading. All images were resized to 224×224×3 and stored in .jpg format. This dataset served as the foundation for training the DenseNet121 recyclability classifier.
Liquid Detection Dataset
This dataset was manually constructed through controlled photography to capture the presence or absence of liquid in waste containers. It consists of 1,056 labeled RGB images across two classes: Liquid and Non-Liquid. Images were collected across a variety of container types and edge cases to reflect the diversity of inputs the prototype bin is expected to encounter.
Multi-Component Detection Dataset
A second self-constructed dataset was created for the task of distinguishing single-component from multi-component waste items. This dataset was built to enable the model to identify items composed of two or more visually distinct components (e.g., a paper cup with a plastic lid), which would otherwise be misclassified by downstream material classifiers.
Real-world validation data was provided by Dawar, an Egyptian waste management company. This dataset contains images of waste collected from within the Egyptian community, enabling the evaluation of the full pipeline's performance on regionally representative, real-world inputs. For the single-object Arabic-labelled dataset, existing class labels were manually mapped to the four material categories. For the unlabeled mixed-waste dataset, object detection using YOLOv8 was employed to isolate individual items, and a labelling classifier (ResNet50, fine-tuned for recyclability) was used to generate labels. After augmentation to balance class distributions, the unified pipeline achieved 92.4% accuracy on this external validation set.
| Dataset | Source | Task | Classes | Size |
|---|---|---|---|---|
| Waste Material Classification | Kaggle | Material Classification | 4 (Paper, Plastic, Organic, Other) | 11,522 images |
| Recyclability | GitHub | Recyclability Classification | 2 (Recyclable, Non-Recyclable) | 34,459 images |
| Liquid Detection | Self-constructed | Liquid Detection | 2 (Liquid, Non-Liquid) | 1,056 images |
| Multi-Component Detection | Self-constructed | Component Detection | 2 (Single, Multi-Component) | 1,380 images |
| Dawar External Validation | Industry Partner | Full Pipeline Validation | 5 (Output Compartments) | Varies |
The mapping from raw categories to pipeline classes is as follows:
| Raw Category | Pipeline Class |
|---|---|
| Paper waste | Paper |
| Plastic waste | Plastic |
| Organic waste | Organic |
| E-waste, automobile, battery, glass, light bulbs, metal | Other |
| Evaluation | Accuracy |
|---|---|
| Material Classification (validation) | 94.3% |
| Recyclability Classification (validation) | 96.8% |
| Liquid Detection (validation) | 94.7% |
| Component Detection (validation) | 94.5% |
| Unified pipeline on external real-world data | 92.4% |
Figure 2. Physical SmartSort prototype developed in collaboration with AUC's Eltoukhy Learning Factory and MakersGate.
Beyond the software pipeline, this project was extended into a real-world smart bin prototype funded by Nestlé and Dawar, and fabricated by AUC's Eltoukhy Learning Factory and MakersGate.
The prototype integrates key hardware and software components to enable fully automated, real-time waste sorting:
- Camera module for image capture at the point of disposal.
- Embedded hardware running the deep learning pipeline for on-device inference.
- Motorized compartments driven by the sequential decision pipeline (Decision Codes 1–7).
- A 2-DOF pan-tilt sorting mechanism to physically route waste to its designated compartment.
- Sensors for bin capacity and status monitoring.
This project would not have been possible without the guidance and unwavering support of:
- Dr. Noha Youssef — Thesis advisor, Department of Mathematics and Actuarial Science, AUC
- Ms. Mahira Hassan — Nestlé, for industry collaboration and funding
- Mr. Amr Fathi & Mr. Youssef Sami — Dawar, for real-world data and deployment partnership
- AUC Eltoukhy Learning Factory & MakersGate — Prototype design and fabrication
This repository is shared for academic and research purposes.