File size: 2,597 Bytes
ef1004f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | ---
license: apache-2.0
library_name: onnx
pipeline_tag: image-segmentation
tags:
- onnx
- onnxruntime-web
- background-removal
- salient-object-detection
- u2net
---
# u2netp (ONNX)
A mirror of **u2netp** — the small ("portable") variant of
[U²-Net](https://github.com/xuebinqin/U-2-Net) — for salient object detection,
the model behind background removal.
This repository exists so the model can be fetched directly by a browser at
runtime. It is a **verbatim mirror**: the file is byte-for-byte the `u2netp.onnx`
asset published by the [rembg](https://github.com/danielgatis/rembg) project, and
its SHA-256 is pinned by the consumer.
| | |
| --- | --- |
| File | `u2netp.onnx` |
| Size | 4,574,861 bytes (4.36 MiB) |
| SHA-256 | `309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8` |
| Opset | 11 |
## Tensor contract
Read the input and output names off the graph rather than hardcoding them — the
first of each is what the pipeline below uses.
- **Input** — `float32[1, 3, 320, 320]`, NCHW, RGB.
- **Output** — `float32[1, 1, 320, 320]`, a saliency map. U²-Net emits several
side outputs; the **first** is the fused one to use.
## Pre / post-processing
The model is fixed at 320×320, so the image is **squashed** to that size (aspect
ratio is not preserved), then normalised with the ImageNet statistics:
```
scaled = (pixel / 255) / max_pixel_value_of_the_image
tensor = (scaled - mean) / std
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
```
Afterwards, min–max normalise the saliency map to 0–1, scale to 0–255, resize it
back to the original dimensions (bilinear), and use it as the alpha channel of
the source image.
> **Note on other mirrors.** Some copies of this model ship a
> `preprocessor_config.json` describing a letterbox resize
> (`keep_aspect_ratio` / `do_pad`). That does **not** match the pipeline above,
> and following it changes the output.
## Licence and credit
The U²-Net weights are **Apache-2.0**. If you use them, cite the paper:
```bibtex
@InProceedings{Qin_2020_PR,
title = {U2-Net: Going Deeper with Nested U-Structure for Salient Object Detection},
author = {Qin, Xuebin and Zhang, Zichen and Huang, Chenyang and Dehghan, Masood
and Zaiane, Osmar and Jagersand, Martin},
journal = {Pattern Recognition},
volume = {106},
pages = {107404},
year = {2020}
}
```
- Model and paper: [xuebinqin/U-2-Net](https://github.com/xuebinqin/U-2-Net) (Apache-2.0)
- The release asset mirrored here: [danielgatis/rembg](https://github.com/danielgatis/rembg) (MIT)
|