DDUNet: Dual Dynamic U-Net for Highly-Efficient Cloud Segmentation

Yijie Li 1 , Hewei Wang 2 , Jinfeng Xu 3 , Puzhen Wu 4 , Yunzhong Xiao 2 , Shaofan Wang 5 , Soumyabrata Dev 4,6 · 2025

Read DDUNet: Dual Dynamic U-Net for Highly-Efficient Cloud Segmentation by Yijie Li 1 , Hewei Wang 2 , Jinfeng Xu 3 , Puzhen Wu 4 , Yunzhong Xiao 2 , Shaofan Wang 5 , Soumyabrata Dev 4,6 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-15385.txt.

Description

Yijie Li and co-authors build a lean neural network that outlines clouds in sky images. Their 2025 paper lists three flaws in older methods: fixed-size filters limit what the network can see, results wobble across scenes, and heavy models cannot run in real time. Their Dual Dynamic U-Net answers with two adaptive modules.

One module mixes features across several scales on the fly, widening the view where needed. The other generates its own weights and biases for the final sorting layers, which helps the model travel across datasets. Thanks to depth-wise tricks, the whole network holds just 0.33 million parameters. It still reaches 95.3 percent accuracy on a standard sky dataset and beats rivals in both accuracy and speed across three test setups.

Catalog record

CREATORYijie Li 1 , Hewei Wang 2 , Jinfeng Xu 3 , Puzhen Wu 4 , Yunzhong Xiao 2 , Shaofan Wang 5 , Soumyabrata Dev 4,6
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE17 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-15385

Browse the full library · About the Tower of Babel