AWS Sheep Recognition Case Study | Abdulrahman Alameer

The Challenge

Saudi Arabia has several distinct local sheep breeds, each with unique physical characteristics. Identifying them accurately requires domain expertise that isn't always available. The goal was to build an accessible web tool that anyone could use — simply upload a photo and get an instant breed classification.

The system needed to:

Accurately distinguish between 4 local Saudi sheep breeds
Provide a simple, user-friendly web interface for photo uploads
Return results with high confidence (≥95%)
Be deployed as a publicly accessible, containerized service

Architecture Overview

The solution uses a fully AWS-native architecture with Docker containerization, managed container orchestration, and a custom-trained ML model.

AWS Architecture Diagram showing the full flow: Client to Route 53 to Internet Gateway to ECS/EC2 in a VPC, with ECR for container images, S3 for storage, and Rekognition for breed analysis.

Fig 1. AWS Architecture — Sheep Breed Recognition System

How It Works

Step 01 — Build

Docker Image on EC2

The web application was Dockerized and the image was built on a private EC2 instance for security, then pushed to Amazon ECR.

Step 02 — Deploy

ECS Public Service

The container runs on Amazon ECS as a public-facing web service, exposed through an Internet Gateway within a VPC.

Step 03 — Route

Route 53 DNS

Amazon Route 53 handles domain name resolution, providing a clean URL for users to access the application.

Step 04 — Analyze

Rekognition Classification

Uploaded photos are stored in S3. Amazon Rekognition (custom-trained model) analyzes the image and returns the breed type to the client.

Implementation

The core classification logic uses the Boto3 SDK to call Amazon Rekognition's custom labels API:

import boto3

def identify_breed(photo):
    client = boto3.client('rekognition')
    response = client.detect_custom_labels(
        Image={'S3Object': {'Bucket': 'sheep-data', 'Name': photo}},
        MinConfidence=95
    )

The MinConfidence=95 threshold ensures only high-confidence predictions are returned, reducing false positives across the 4 breed categories.

Tech Stack

EC2 ECR ECS S3 Rekognition Route 53 Docker Python

Infrastructure Details

VPC Design: Public subnet hosts the ECS service (Internet Gateway access); private subnet hosts the EC2 build instance for security isolation
Container Registry: Docker images stored in Amazon ECR, pulled by ECS task definitions at deployment
Storage: User-uploaded images land in an S3 bucket, which Rekognition reads directly via S3Object references
DNS: Route 53 provides a human-readable domain pointing to the ECS service endpoint
Security: Build process isolated on private EC2; ECS service exposed only through controlled Internet Gateway path

Key Takeaways

Custom-trained Rekognition models can achieve production-grade accuracy for niche classification tasks
Containerizing the application with Docker + ECS provides scalable, reproducible deployments
Separating the build environment (private EC2) from the runtime (public ECS) follows security best practices
The S3 → Rekognition pipeline eliminates the need to transfer image data through the application layer

Saudi Sheep Breed Identifier

Context

Role

Stack

Outcome