ISCC - Semantic Image-Code

iscc-sci is a proof of concept implementation of a semantic Image-Code for the ISCC (International Standard Content Code). Semantic Image-Codes are designed to capture and represent the semantic content of images for improved similarity detection.

Caution

This is a proof of concept. All releases with version numbers below v1.0.0 may break backward compatibility and produce incompatible Semantic Image-Codes. The algorithms of this iscc-sci repository are experimental and not part of the official ISO 24138:2024 standard.

What is ISCC Semantic Image-Code

The ISCC framework already comes with an Image-Code that is based on perceptual hashing and can match near duplicates. The ISCC Semantic Image-Code is planned as a new additional ISCC-UNIT focused on capturing a more abstract and broad semantic similarity. As such the Semantic Image-Code is engineered to be robust against a broader range of variations that cannot be matched with the perceptual Image-Code.

Features

• Semantic Similarity: Leverages deep learning models to generate codes that reflect the semantic content of images.
• Bit-Length Flexibility: Supports generating codes of various bit lengths (up to 256 bits), allowing for adjustable granularity in similarity detection.
• ISCC Compatible: Generates codes that are fully compatible with the ISCC specification, facilitating integration with existing ISCC-based systems.

Installation

Ensure you have Python 3.11 or newer installed on your system. The package requires an ONNX runtime that is selected via install extras. For CPU inference (works everywhere):

pip install "iscc-sci[cpu]"

For NVIDIA CUDA accelerated inference (requires CUDA 12.x and cuDNN 9.x):

pip install "iscc-sci[gpu]"

Note

Install exactly one of the cpu/gpu extras. The underlying onnxruntime and onnxruntime-gpu packages unpack into the same directory and overwrite each other, so installing both silently disables GPU support. A plain pip install iscc-sci installs no ONNX runtime and fails on import with instructions.

Usage

To generate a Semantic Image-Code for an image, use the code_image_semantic function. You can specify the bit length of the code to control the level of granularity in the semantic representation.

import iscc_sci as sci

# Generate a 64-bit ISCC Semantic Image-Code for an image file
image_file_path = "path/to/your/image.jpg"
semantic_code = sci.code_image_semantic(image_file_path, bits=64)

print(semantic_code)

How It Works

iscc-sci uses a pre-trained deep learning model based on the 1st Place Solution of the Image Similarity Challenge (ISC21) to create semantic embeddings of images. The model generates a feature vector that captures the essential characteristics of the image. This vector is then binarized to produce a Semantic Image-Code that is robust to variations in image presentation but sensitive to content differences.

Model

The model file is downloaded on first use from the iscc-binaries v1.0.0 release and verified against its BLAKE3 checksum:

• File: iscc-sci-v0.1.0-w16.onnx, 104,790,555 bytes
• BLAKE3: 9f24dd3d0440eaa07adf121b21a1cbd8518fe36048b9d8a72b606d8a6416e8a3
• Input: 1×3×512×512 RGB, normalized as (x / 255 - 0.5) / 0.5
• Output: 256 floats. The Semantic-Code is made of their sign bits.

Provenance:

1. Backbone: EfficientNetV2-M, trained by Google on ImageNet-21k and fine-tuned on ImageNet-1k (timm/tf_efficientnetv2_m.in21k_ft_in1k, Apache-2.0).
2. Training: fine-tuned on the 1M DISC21 training images by Shuhei Yokoo (lyakaap) for the Facebook AI Image Similarity Challenge 2021, descriptor track, 1st place (ISC21-Descriptor-Track-1st, report arXiv 2112.04323). The weights are MIT licensed, see LICENSE-MODEL.
3. Export: converted to ONNX with the image-embedding/isc operator of Towhee (Apache-2.0).
4. Compression: weights stored as fp16 and cast back to fp32 when the model loads (all maths stays fp32), graph simplified with onnxslim, input fixed to 1×3×512×512. This is the same file the ISCC C2PA demo uses.

Licence: the iscc-sci code is Apache-2.0 and the model weights are MIT. The DISC21 training data is governed by Meta's Image Similarity Dataset License Agreement in the competition rules. Its clause 2 covers models trained on the data: "Such algorithms, software, machine learning models, techniques and technologies may be used for academic and commercial purposes." Commercial use is allowed.

The "CC BY-NC 4.0" licence that is sometimes cited for the model belongs to Meta's evaluation code repository facebookresearch/isc2021, not to the dataset.

Development

This is a proof of concept and welcomes contributions to enhance its capabilities, efficiency, and compatibility with the broader ISCC ecosystem. For development, install the project with uv. The default uv sync installs the test group, which provides a CPU ONNX runtime:

git clone https://github.com/iscc/iscc-sci.git
cd iscc-sci
uv sync

Contributing

Contributions are welcome! If you have suggestions for improvements or bug fixes, please open an issue or pull request. For major changes, please open an issue first to discuss what you would like to change.