This repository contains the final project report for Cloud Computing (2024/2025) at Sapienza Università di Roma.
The goal of this project is to implement a fully serverless, cloud-based image preprocessing pipeline using AWS services. The pipeline automatically triggers an image processing function whenever a new image is uploaded to Amazon S3. The images are converted to grayscale and resized using the Pillow library in Python, with the processed outputs stored in a separate S3 bucket.
This system is optimized for scalability, cost-efficiency, and real-time operation using a completely serverless architecture.
- 📥 Automatic Triggering: Uploading an image to the input S3 bucket triggers processing automatically.
- 🖼️ Image Processing: Grayscale conversion and resizing to 256×256 using the Pillow library.
- 📤 Processed Output: Results are saved to a separate S3 bucket.
- ⚙️ Serverless Design: No servers or containers required—powered entirely by AWS Lambda.
- 📊 Monitoring: CloudWatch tracks metrics like function duration, error rate, and concurrency.
- 🔐 Secure Access: Managed through AWS IAM roles and permissions.
| Technology | Purpose |
|---|---|
| AWS Lambda | Executes the image processing logic in response to uploads |
| Amazon S3 | Stores input and output images and triggers Lambda |
| Amazon CloudWatch | Observability, metrics, and debugging |
| AWS IAM | Secures access to S3 and Lambda |
| Python + Pillow | Performs image transformations |
[User Uploads Image]
│
▼
┌────────────────────┐
│ Amazon S3 (Input) │
└────────────────────┘
│
▼ (S3:ObjectCreated trigger)
┌────────────────────┐
│ AWS Lambda │
│ - Convert to Gray │
│ - Resize to 256x256│
└────────────────────┘
│
▼
┌────────────────────┐
│ Amazon S3 (Output) │
└────────────────────┘
- Runtime: Python 3.12
- Memory: 128 MB
- Timeout: 3 seconds
- Trigger:
s3:ObjectCreated:*event - Environment Variable:
OUTPUT_BUCKETspecifies the output S3 bucket - Pillow Library: Included via a Lambda Layer
Performance was evaluated under two scenarios:
Images were uploaded manually via CLI or browser in batches of:
- 10
- 50
- 100
- 500
- 1000
Simulated high-throughput uploads using a Python script to rapidly copy images into S3.
- Invocation Count
- Execution Duration
- Success Rate & Error Count
- Throttling Events
- Concurrent Executions
- Async Event Age
- Events Received/Dropped
- The system scaled efficiently with image volume and resolution.
- No throttling occurred, even under high-rate uploads.
- Execution duration increased with resolution but remained within acceptable limits.
- Brief cold start delays were observed and recovered quickly.
- Both trigger types demonstrated resilience and reliability.
| Resolution | Input (RGB) | Output (Grayscale) |
|---|---|---|
| Low | 1200 x 800 pixels | 256 x 256 pixels |
| Mid | 1500 x 997 pixels | 256 x 256 pixels |
| High | 2040 x 1536 pixels | 256 x 256 pixels |
All graphs and performance metrics collected from CloudWatch are organized and available here:
📦 Google Drive – Metrics & Graphs
Includes:
- Manual Trigger graphs
- High-Rate Trigger graphs
- All image resolutions (Low, Mid, High)
- All batch sizes (10–1000)
This project successfully demonstrates the use of AWS services to build a serverless image preprocessing pipeline. The system is lightweight, scalable, and designed for real-time event-driven workloads. It offers:
- No server management
- Automatic scaling
- Minimal cost (pay-per-use)
- Consistent performance across workloads
This architecture can be extended to:
- Support additional image formats
- Integrate ML-based image classification
- Build a front-end with API Gateway for end users
- Beyza Nur Elaslan (2180876)
- Marthe Elgawly (2170201)
- Murad Hüseynov (2181584)
📅 Final Project – Cloud Computing (2024/2025)
🏛️ Sapienza Università di Roma
This repository is academic, but feedback and forks are welcome for educational or extension purposes.