Object Retrieval System for Image

This repository contains an implementation of an image retrieval system designed to search for and rank similar object in images from a large gallery based on query images. The system leverages CNN-based feature extraction to enhance retrieval performance. It is built to process a gallery of images and return ranked results based on similarity to a given query.

Example

Features

1. A CNN-based approach using a pre-trained ResNet50 model for feature extraction, paired with cosine similarity for ranking.

• Modular Design: Code is organized into separate folders for each method, making it easy to extend or modify.

Repository Structure

• extract_feature_multi_process.py: Feature extraction using ResNet50.
• all_retrieval.py: Image retrieval and ranking using cosine similarity.

Methodology

Method 1: CNN-Based Retrieval (ResNet50)

This method extracts deep features from images using a pre-trained ResNet50 model and ranks gallery images based on cosine similarity.

• Feature Extraction:
  • Images are pre-processed (BGR to RGB conversion) using OpenCV.
  • Objects are cropped using a Fast R-CNN model when multiple instances are present.
  • ResNet50 extracts features after removing the fully connected layer, followed by global average pooling to generate a 1D feature vector.
• Retrieval:
  • Cosine similarity is computed between the query feature vector and gallery feature vectors.
  • Top 10 images with the highest similarity scores are returned.

Setup and Usage

Prerequisites

• Python 3.x
• Libraries: OpenCV, NumPy, PyTorch (for ResNet50), and other dependencies (see requirements.txt if provided).
• A gallery of images (not included in this repo).

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/image-retrieval-system.git
   cd image-retrieval-system

2. Install dependencies:

   pip install -r requirements.txt

3. Add your gallery images to the gallery/ folder and query images to the query/ folder.

Running the Code

• For CNN-based retrieval:

  python /extract_feature_multi_process.py
  python /all_retrieval.py

Notes

• The gallery/ folder is empty to save space. Populate it with your image dataset before running the code.
• Query images and bounding box annotations (if required) should be placed in the query/ folder.
• Adjust file paths in the scripts if your dataset structure differs.