项目作者: apennisi

项目描述 :
Attention Squeeze U-Net
高级语言: Python
项目地址: git://github.com/apennisi/att_squeeze_unet.git
创建时间: 2021-04-06T06:06:34Z
项目社区:https://github.com/apennisi/att_squeeze_unet

开源协议:

下载


Attention Squeeze U-Net

Attention Squeeze U-Net is a network for Skin Lesion Segmentation able to run on embedded devices.

The network is made as shown in the following image:

Attention U-Net

The network has been trained on

Requirements

  • Numpy
  • Tensorflow (>= 2.0)
  • Pillow
  • OpenCV (>= 4.0)

Usage

You can download isic2017 from this link.
If you want to use your own dataset, the training, validation, and testing folder have to contain the the images in jpg format and the masks in png format.

Training

  1. python train.py --train_set /path/to/isic2017 --eval_set /path/to/validation/isic2017 --checkpoint_dir /path/to/directory/where/you/want/to/save/the/models --aug_scale 3 --network attention_squeeze_unet (or squeeze_unet, attention_unet, unet, segnet)

Testing

  1. python test.py --test_dir /path/to/testing/isic2017 --resume /path/to/the/trained/model --network attention_squeeze_unet (or squeeze_unet, attention_unet, unet, segnet)

Inference

  1. python test.py --test_dir /path/to/testing/isic2017 --resume /path/to/the/trained/model --save_dir /path/to/the/directory/to/save/masks --network attention_squeeze_unet (or squeeze_unet, attention_unet, unet, segnet)

Output Example

Attention Squeeze U-Net has only 2.6 millions of parameters to train. The output of the inference proces is as follows:

Output Attention Squeeze U-Net

More Info and Citation:

  1. @article{Pennisi2022SkinLA,
  2. title={Skin Lesion Area Segmentation Using Attention Squeeze U-Net for Embedded Devices},
  3. author={Andrea Pennisi and Domenico Daniele Bloisi and Vincenzo Suriani and Daniele Nardi and Antonio Facchiano and Annalisa Giampetruzzi},
  4. journal={Journal of Digital Imaging},
  5. year={2022},
  6. volume={35},
  7. pages={1217 - 1230}
  8. }