U-2-Net

Model Description

U-2-Net is a deep learning model designed for image segmentation tasks, particularly for generating detailed masks. It leverages a unique architecture with nested U-blocks that allow the model to capture both high-level semantic features and fine-grained details. U-2-Net has demonstrated high performance in a variety of segmentation tasks, making it a versatile choice for applications such as background removal, object detection, and medical image analysis.

Usage

Perform mask generation with BritishWerewolf/U-2-Net.

Example

import { AutoModel, AutoProcessor, RawImage } from '@huggingface/transformers';

const img_url = 'https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png';
const image = await RawImage.read(img_url);

const processor = await AutoProcessor.from_pretrained('BritishWerewolf/U-2-Net');
const processed = await processor(image);

const model = await AutoModel.from_pretrained('BritishWerewolf/U-2-Net', {
    dtype: 'fp32',
});

const output = await model({ input: processed.pixel_values });
// {
//   mask: Tensor {
//     dims: [ 1, 320, 320 ],
//     type: 'uint8',
//     data: Uint8Array(102400) [ ... ],
//     size: 102400
//   }
// }

Model Architecture

The U-2-Net model is built upon a nested U-structure, where each U-block consists of multiple convolutional layers, pooling, and up-sampling operations. The architecture features a combination of down-sampling and up-sampling paths, enabling the model to learn features at different scales. This design allows the U-2-Net to produce accurate and high-resolution segmentation maps. The key components of the architecture include Residual U-blocks (RSU) that enhance feature representation and ensure efficient information flow through the network.

Inference

To use the model for inference, you can follow the example provided above. The AutoProcessor and AutoModel classes from the transformers library make it easy to load the model and processor.

Credits

• rembg for the ONNX model.
• The authors of the original U-2-Net model can be credited at https://github.com/xuebinqin/U-2-Net.

Licence

This model is licensed under the Apache License 2.0 to match the original U-2-Net model.