Image Recommender

Overview

This project is designed to help you find the top five images similar to a given image from a large dataset of nearly 450,000 images. By leveraging Python, this software can recommend images based on various similarity metrics, such as color histograms and embeddings. The project focuses on image processing, feature extraction, similarity calculation, and clustering.

Getting Started

Prerequisites Before you begin, ensure that Python is installed on your system. We recommend using Python 3.8 or higher for optimal compatibility.

Installation

Clone the Repository

Begin by cloning the repository to your local machine:

git clone https://github.com/yourusername/image-recommender.git

Navigate into the Project Directory and install the dependencies.

The necessary dependencies are:

numpy==1.24.3 pandas==2.0.3 tqdm==4.66.1 matplotlib==3.7.1 Pillow==9.4.0 opencv-python==4.7.0.72 scipy==1.11.2 sqlite==3.41.2 torch==2.0.1 torchvision==0.15.2 scikit-learn==1.3.0 umap-learn==0.5.4 plotly==5.15.0 pytest==8.3.2

Setting Up the Dataset

Before you can start recommending images, you'll need to set up your dataset with all the necessary similarities and metadata.

Load and Process Images

In your main Python script or Jupyter notebook, use the following code to load and process images:

batch_size = 1000
desired_size = (224, 224)
main_load_images(batch_size, desired_size)

This script will process the images from your specified directory, calculate the relevant features, and save them as pickle files for later use.

Recommending Images

Load the Pickled Data Make sure your data is properly loaded from the pickle files:

initialize_data()

Find Similar Images

To find a specified number of similar images based on your chosen feature type and distance metric, use the following command:

main_finding_similarities(1, "RGB", "euclidean", 5)

This function will compute the similarities between your input images and the dataset, displaying the top n similar images. You can specify the number of input images, choose between 'RGB', 'HSV' or 'Embedding' as the feature type and select 'euclidean', 'manhattan' or 'cosine' as the distance metric. Important: You need to specify the paths to your input images before running the function.

Correct Data Inconsistencies

If you encounter any inconsistencies in the data, you can correct them by running:

correct_data()

Directory Structure

• resources/: Contains all the Python modules for image processing, feature extraction, and similarity calculations.
• main.ipynb: The main Jupyter notebook where functions are called, and processes are executed.

Contact

For any questions or further information, feel free to contact:

• Tim Sandrock: tim.sandrock@study.hs-duesseldorf.de
• Joschua Schramm: joschua.schramm@study.hs-duesseldorf.de