Garbage Classification API

A CNN-powered REST API that identifies recyclable waste categories from photos in real time — enabling automated sorting decisions at recycling facilities and smart bin deployments.

---

Business Problem

Manual waste sorting is one of the largest operational costs in recycling infrastructure — misclassification rates above 20% cause contamination of recyclable streams, leading to material rejection and landfill overflow. An automated vision system that correctly identifies cardboard, glass, metal, paper, plastic, and general trash from a single photo enables real-time conveyor belt sorting, reduces human labor costs, and directly improves recycling yield rates for facilities processing thousands of items per hour.

---

Demo

Upload an image through the Streamlit interface and get an instant classification result.

Supported waste categories: cardboard · glass · metal · paper · plastic · trash

---

Results

Metric	Score
Train Accuracy	77.44%
Test Accuracy	75.3%

Model: Custom 3-block CNN (Conv2d → BatchNorm2d → ReLU → MaxPool2d), trained from scratch, no transfer learning.

Baseline (random classifier, 6 classes): Accuracy = 16.7% ↑ +58.6 pp improvement vs baseline

---

Dataset

• Source: Garbage Classification (Kaggle) — asdasdasasdas/garbage-classification
• Size: ~2,527 real-world waste images across 6 categories
• Split: 80% train / 20% test via random_split (seed=42)
• Preprocessing: resized to 128×128 for both training and inference
• Class balance: near-balanced, no resampling required

---

Approach

1. Data Loading — downloaded via kagglehub, loaded with torchvision.datasets.ImageFolder; 80/20 train/test split with fixed seed
2. Augmentation (train only) — RandomHorizontalFlip, RandomRotation(15), ColorJitter(brightness=0.3, contrast=0.3) to improve robustness to real-world photo variability (lighting, angle)
3. Normalization — Normalize([0.5]*3, [0.5]*3), identical for train/test
4. Model Architecture — 3-block CNN: Conv2d(3→32→64→128) + BatchNorm2d
   • ReLU + MaxPool2d(2) per block; classifier head with Dropout(0.4) and Linear(128*16*16 → 256 → 6)
5. Training — 30 epochs, Adam (lr=0.0003), CrossEntropyLoss, loss reduced from ~123 to ~54 over training
6. Inference App — Streamlit interface with image upload, cached model loading (@st.cache_resource), CUDA/CPU auto-detection

---

Key Challenges & Solutions

Small real-world dataset (~2,500 images) prone to overfitting With ~400 images per class and significant real-world photo variability (lighting, background, angle), a plain CNN overfits quickly → added RandomHorizontalFlip, RandomRotation(15), ColorJitter on the training split and Dropout(0.4) in the classifier head → train/test gap stayed under 3 percentage points (77.44% vs 75.3%), indicating the model generalizes rather than memorizes.

Consistent train/inference preprocessing Mismatched resize/normalization between the training notebook and the production app is a common source of silent accuracy drops → both GarbageClassification.ipynb and main.py use identical Resize((128,128))

• Normalize([0.5]*3, [0.5]*3), and the same GarbageClassifier class definition is duplicated exactly in both files → guarantees the served model's feature map dimensions match the trained checkpoint.

Uploaded images in arbitrary formats (PNG/RGBA, JPEG) Raw uploads can arrive as RGBA, grayscale, or other formats incompatible with the 3-channel input the model expects → Image.open(...).convert('RGB') applied unconditionally before the transform pipeline → all supported formats are normalized to a valid model input, with errors caught and surfaced via st.error() instead of crashing the app.

---

Tech Stack

Category	Tools
Language	Python 3.11
ML	PyTorch, torchvision
App	Streamlit
Data	KaggleHub, Pillow, Matplotlib
Regularization	BatchNorm2d, Dropout

---

Project Structure

GarbageClassification/
├── .gitignore
├── readme.md
├── requirements.txt
└── GarbageClassification/
    ├── GarbageClassification.ipynb
    ├── datasets/
    │   └── trash_dataset.zip
    ├── labels_GarbageClassification.pth
    ├── main.py
    ├── model_GarbageClassification.pth
    ├── sample_images.png
    └── tests/
        ├── cardboard117.jpg
        ├── glass11.jpg
        ├── metal103.jpg
        ├── paper104.jpg
        ├── plastic103.jpg
        └── trash113.jpg

---

How to Run

# 1. Clone and install
git clone https://github.com/your-username/garbage-classification
cd garbage-classification/GarbageClassification
pip install torch torchvision streamlit pillow

# 2. Train the model (optional — pretrained weights included)
# open and run GarbageClassification.ipynb (Colab-ready, GPU recommended)

# 3. Launch the app
streamlit run main.py
# Opens at http://localhost:8501

---

Business Impact

• ↓ reduction in manual sorting labor costs at recycling facilities vs fully manual classification pipelines (estimated)
• ↑ 75.3% automated classification accuracy across 6 material types, trained from scratch on a compact real-world dataset with no external pretrained weights
• ↑ Streamlit interface enables non-technical staff to validate sorting decisions without API integration overhead
• ↑ Retrainable on proprietary facility-specific waste categories with minimal code changes — no vendor lock-in

---