Reverse Image Search

This repository contains my undergraduate thesis project developed at the Faculty of Informatics in Pula, focusing on reverse image search using the power of Milvus.io and Towhee.io.

Author

Luka Blašković (luka.blaskovic@student.unipu.hr)

Short description

Vector databases are used to store unstructured data such as images, videos, music, sensor data, and so on. Before importing the data into the Milvus database, it first has to be encoded into vector embeddings using towhee.io. An “embedding” vector is a numeric representation of unstructured data, images in this case. Embedding vectors are then loaded into Milvus database. Milvus supports different types of similarity metrics and indexes for different use cases. Here, we will use L2 euclidian distance metric and IVF_FLAT index. Application can be tested using Gradio web interface which runs from build.ipynb file.

Undergraduate thesis: Available Here

Tools used

• Python 3.8.10
• Milvus.io
• Towhee.io
• Jupyter notebook

Dependencies

! python -m pip install -q pymilvus towhee gradio opencv-python pillow pyarrow

Dataset

curl -L https://github.com/towhee-io/examples/releases/download/data/reverse_image_search.zip -O

Organization

Juraj Dobrila University of Pula
Pula Faculty of Informatics
Undergraduate thesis - Vector databases for unstructured data processing, 2021./2022.
Course: Databases II
Mentor: doc. dr. sc. Goran Oreški (https://fipu.unipu.hr/fipu/goran.oreski, goran.oreski@unipu.hr)