[中文] | [English]

Coin-CLIP 🪙 : Enhancing Coin Image Retrieval with CLIP

Coin-CLIP breezedeus/coin-clip-vit-base-patch32 is built upon OpenAI's CLIP (ViT-B/32) model and fine-tuned on a dataset of more than 340,000 coin images using contrastive learning techniques. This specialized model is designed to significantly improve feature extraction for coin images, leading to more accurate image-based search capabilities. Coin-CLIP combines the power of Visual Transformer (ViT) with CLIP's multimodal learning capabilities, specifically tailored for the numismatic domain.

Key Features:

State-of-the-art coin image retrieval;
Enhanced feature extraction for numismatic images;
Seamless integration with CLIP's multimodal learning.

To further simplify the use of the Coin-CLIP model, this project provides tools for quickly building a coin image retrieval engine.

Comparison: Coin-CLIP vs. CLIP

Example 1 (Left: Coin-CLIP; Right: CLIP)

Example 2 (Left: Coin-CLIP; Right: CLIP)

More Examples

Example 3 (Left: Coin-CLIP; Right: CLIP)

Example 4 (Left: Coin-CLIP; Right: CLIP)

Example 5 (Left: Coin-CLIP; Right: CLIP)

Example 6 (Left: Coin-CLIP; Right: CLIP)

Install

pip install coin_clip

Usage

Code Examples

Extract Feature Vectors from Coin Images

from coin_clip import CoinClip

# Automatically download the model from Huggingface
model = CoinClip(model_name='breezedeus/coin-clip-vit-base-patch32')
images = ['examples/10_back.jpg', 'examples/16_back.jpg']
img_feats, success_ids = model.get_image_features(images)
print(img_feats.shape)  # --> (2, 512)

⚠️ Note:

The above code automatically downloads the breezedeus/coin-clip-vit-base-patch32 model from Huggingface. If you cannot download automatically, please manually download the model locally, and then initialize CoinClip by specifying the local directory of the model through the model_name parameter, like model_name='path/to/coin-clip-vit-base-patch32'.

Command line tools

Building a Vector Retrieval Engine

coin-clip build-db can be used to build a vector search engine. It extracts features from all coin images 🪙 in a specified directory and builds a ChromaDB vector search engine.

$ coin-clip build-db -h
Usage: coin-clip build-db [OPTIONS]

  Extract vectors from a candidate image set and build a search engine based
  on it.

Options:
  -m, --model-name TEXT       Model Name; either local path or huggingface
                              model name  [default: breezedeus/coin-clip-vit-
                              base-patch32]
  -d, --device TEXT           ['cpu', 'cuda']; Either 'cpu' or 'gpu', or
                              specify a specific GPU like 'cuda:0'. Default is
                              'cpu'.  [default: cpu]
  -i, --input-image-dir TEXT  Folder with Coin Images to be indexed. [required]
  -o, --output-db-dir TEXT    Folder where the built search engine is stored.
                              [default: ./coin_clip_chroma.db]
  -h, --help                  Show this message and exit.

For instance,

$ coin-clip build-db -i examples -o coin_clip_chroma.db

Querying

After building the vector search engine with the above command, you can use the coin-clip retrieve command to retrieve the coin images 🪙 most similar to a specified coin image.

$ coin-clip retrieve -h
Usage: coin-clip retrieve [OPTIONS]

  Retrieve images from the search engine, based on the query image.

Options:
  -m, --model-name TEXT  Model Name; either local path or huggingface model
                         name  [default: breezedeus/coin-clip-vit-base-
                         patch32]
  -d, --device TEXT      ['cpu', 'cuda']; Either 'cpu' or 'gpu', or specify a
                         specific GPU like 'cuda:0'. Default is 'cpu'.
                         [default: cpu]
  --db-dir TEXT          Folder where the built search engine is stored.
                         [default: ./coin_clip_chroma.db]
  -i, --image-fp TEXT    Image Path to retrieve  [required]
  -h, --help             Show this message and exit.