angganurfaizal / similar-ordinals

Finding similar ordinals pictures

Ordinals pictures similarity

This project focuses on creating a service that displays similar ordinal pictures.

Its UI is available at https://ordsimilarity.com/. The API (details below) is available at https://api.ordsimilarity.com.

The pictures are classified using the average hash algorithm, which converts them into a bit sequence of 0s and 1s. They are then compared using the Hamming distance, which measures how many bits are common between two pictures. A detailed description can be found at this link.

The average hashes of all ordinals are stored in a JSON file. Example data can be seen in average_hash_example.json.

NOTE: we use JSON file instead of a "normal DB" because we need to always loop through all the data inside, and it turns out it is much faster to load all data from JSON file than from some DB like sqlite3.

The main processing function is get_matches_from_data in get_matches.py. This function iterates through all the given average hashes and compares their similarity with the average hash calculated from the provided data - either an ordinal ID or the custom picture content. It returns a list of most similar ordinals, sorted by their similarity.

get_matches.py also provides CLI access to the function. See python get_matches.py --help for more details.

API

The API is implemented in python using the FastAPI framework. On startup, it loads the average hashes from the JSON file into memory. Then, for each request, it provides this data to the above-mentioned get_matches_from_data function and collects the result. Before returning the JSON result to the client, it enriches the similar ordinals data with additional useful properties or links.

NOTE: The API actually calls the Rust API, which provides much better performance. See the Rust server section for more details.

The API includes multiple endpoints for similarity searches, which are defined in api.py:

• GET /ord_id/{ord_id}?top_n=N
  • Returns the top N similar ordinal pictures for the given ordinal picture ID.
  • The maximum value for top_n is 20, and it is optional. If not specified, it defaults to 20.
• POST /file?top_n=N
  • Returns the top N similar ordinal pictures for the uploaded ordinal picture, which should be included as a "file" form-data argument.
  • Example usage: curl -X POST -H "Content-Type: multipart/form-data" -F "file=@images/1.jpg" http://localhost:8001/file?top_n=10
  • top_n is unlimited in this case, and is optional. If not specified, it defaults to 20.

Rust bin/lib

For improved performance, the get_matches functionality has also been implemented in Rust, to speed up the similarity search.

similar_pictures contains a Rust binary and library for this purpose.

The connection to Rust from python is established in get_matches_rust.py, which exposes the same CLI interface as get_matches.py. This connection is made possible by a Rust shared library, residing under similar_pictures/target/release/libsimilar_pictures.so.

Additionally, a standalone Rust binary acting as CLI is created, under similar_pictures/target/release/similar_pictures. CLI help can be seen by running ./similar_pictures/target/release/similar_pictures --help. Its usage is similar to the python version.

Both the Rust library and binary utilize the common get_matches function from similar_pictures/src/get_matches.rs, which performs the same task as its python version. To build the Rust binary and library for your operating system, run cargo build --release.

Rust server

The approach with loading the average hashes into memory on server startup to serve requests much faster was also replicated in Rust - in rust_http_server.

It contains a (nonpublic) server with similar endpoints as the python API - the only difference is that it does not accept file object, it needs to be given a file_hash directly.

It should not be used directly, but rather as a backend for the python API. The connection is being established in rust_server.py.