Visual Search Engine

An AI-powered visual search application that finds similar images using deep learning embeddings and semantic similarity.

Built by: Mayank Bansal, Arnav Godara

---

Overview

Visual Search Engine is a web application that allows users to upload an image and instantly discover visually similar images from a database. Powered by ResNet50 neural networks and FAISS similarity search, it provides fast and accurate results without relying on text-based queries.

Key Features

• Drag & Drop Interface - Intuitive, user-friendly design
• Real-time Search - Get results in seconds using efficient FAISS indexing
• Deep Learning - ResNet50 pre-trained model for feature extraction
• Semantic Similarity - Find images based on visual content, not metadata
• Responsive Design - Works seamlessly across devices

---

Tech Stack

Component	Technology
Backend	FastAPI (Python)
ML Model	ResNet50 (PyTorch)
Search Engine	FAISS (Facebook AI Similarity Search)
Frontend	HTML, CSS, JavaScript
Database	Pre-computed embeddings (NumPy)

---

Project Structure

Visual Search Engine/
├── app.py                 # FastAPI server & API endpoints
├── index.html             # Frontend interface
├── script.js              # Client-side logic
├── style.css              # UI styling
├── dataset/
│   └── images/            # Database of 1000+ images
├── embeddings.npy         # Pre-computed image embeddings
├── faiss_index.bin        # FAISS index for similarity search
└── image_paths.pkl        # Image file paths mapping

---

How It Works

1. Image Upload → User uploads an image via the web interface
2. Feature Extraction → ResNet50 generates a 2048-dimensional embedding
3. Similarity Search → FAISS finds the 5 most similar images in the index
4. Results Display → Matching images are displayed in the UI

---

Getting Started

Prerequisites

• Python 3.7+
• Required packages (see below)

Installation

1. Clone or Download this repository

2. Install Dependencies

   pip install fastapi uvicorn torch torchvision pillow numpy faiss-cpu

3. Run the Server

   python app.py

   The server will start at http://127.0.0.1:8000

4. Access the Application

   • Open index.html in your browser or navigate to the server URL

---

API Endpoints

GET /

Returns API status

Response: {"message": "API is running"}

POST /search

Search for similar images

• Input: Image file (JPG, PNG, WEBP)
• Output: JSON with 5 most similar image URLs

{
  "results": [
    "http://127.0.0.1:8000/images/image1.jpg",
    "http://127.0.0.1:8000/images/image2.jpg",
    ...
  ]
}

---

Features Explained

ResNet50 Model

Pre-trained deep neural network that extracts high-level visual features (2048 dimensions) from images, capturing semantic meaning.

FAISS Indexing

Efficiently searches millions of vectors to find the most similar embeddings using optimized approximate nearest neighbor search.

CORS Support

Enabled for cross-origin requests, allowing the frontend to communicate with the backend seamlessly.

---

Usage Example

1. Open the application in your browser
2. Click the upload box or drag & drop an image
3. Wait for processing (typically 1-2 seconds)
4. View the 5 most visually similar images from the database

---

Future Enhancements

• Support for different ML models (VGG16, EfficientNet, Vision Transformers)
• Database management UI for adding/removing images
• Advanced filtering and sorting options
• User authentication and saved searches
• Performance optimization for larger datasets
• Mobile app version

---

License

This project is for educational and research purposes.

---

Last Updated: 2026