MCAT-UNet

MCAT-UNet: Convolutional and Cross-shaped Window Attention Enhanced UNet for Efficient High-resolution Remote Sensing Image Segmentation

Introduction

MCAT-UNet is an open-source semantic segmentation model based on mmsegmetation, which mainly focuses on developing advanced remote sensing image segmentation. Article download link https://ieeexplore.ieee.org/abstract/document/10521698

The proposed MCAT-UNet can extract local representations and capture long-range spatial dependencies to segment geographic objects more efficiently in complex scenarios with low computational complexity. In particular, MCAT-UNet achieves more complete predictions for large-scale varied objects and small discrete multiscale objects, where the boundaries remain accurate and smooth.

Install

• First, you need to download mmsegmentation and install it on your server.
• Second, place backbone.py and csheadunet.py in the corresponding directory of mmsegmentation.
• Third, train according to the training strategy of mmsegmentation and the training parameters in our paper.

Pretrained Weights of Backbones

pretrain

Data Preprocessing

Download the datasets from the official website and split them yourself.

Potsdam and Vaihingen Potsdam and Vaihingen

LoveDA LoveDA

Training

You can refer to mmsegmentation document (https://mmsegmentation.readthedocs.io/en/latest/index.html).

Results and Logs for MACT-UNet

LoveDA/Potsdam/Vaihingen

Dataset	Crop Size	Lr Schd	mIoU	#params(Mb)	FLOPs(Gbps)	config	log
Potsdam	512x512	100K	75.44	23.2	18.5	config	github
Vaihingen	512x512	100K	74.52	23.2	18.5	config	github
LoveDa	512x512	100K	53.58	23.2	18.5	config	github

Inference on High-resolution remote sensing image

Cite MCAT-UNet

If you find this work useful or interesting, please consider citing the following BibTeX entry.

@article{wang2024mcat,
  title={MCAT-UNet: Convolutional and Cross-shaped Window Attention Enhanced UNet for Efficient High-resolution Remote Sensing Image Segmentation},
  author={Wang, Tao and Xu, Chao and Liu, Bin and Yang, Guang and Zhang, Erlei and Niu, Dangdang and Zhang, Hongming},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year={2024},
  publisher={IEEE}
}

Acknowledgement

Many thanks the following projects's contributions to MACT-UNet.

• mmsegmentation