Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yijie Li, Hewei Wang, Jinfeng Xu, Puzhen Wu, Yunzhong Xiao, Shaofan Wang, Soumyabrata Dev

TL;DR

DDUNet addresses cloud segmentation challenges by combining a U-Net backbone with Dynamic Multi-scale Convolution (DMSC) for adaptive multi-scale feature fusion and Dynamic Weights & Bias Generator (DWBG) for input-specific classifier parameters, all implemented with depthwise convolutions for efficiency. The approach delivers a lightweight model that achieves $95.3\%$ accuracy with only $0.33$M parameters on SWINySEG, outperforming several baselines in both accuracy and latency. Across day-time and night-time configurations, the method demonstrates robust performance and strong generalization, making it suitable for real-time deployment on resource-constrained devices. The work highlights how dynamic, parameter-efficient modules can significantly boost cloud segmentation performance without sacrificing speed.

Abstract

Cloud segmentation amounts to separating cloud pixels from non-cloud pixels in an image. Current deep learning methods for cloud segmentation suffer from three issues. (a) Constrain on their receptive field due to the fixed size of the convolution kernel. (b) Lack of robustness towards different scenarios. (c) Requirement of a large number of parameters and limitations for real-time implementation. To address these issues, we propose a Dual Dynamic U-Net (DDUNet) for supervised cloud segmentation. The DDUNet adheres to a U-Net architecture and integrates two crucial modules: the dynamic multi-scale convolution (DMSC), improving merging features under different reception fields, and the dynamic weights and bias generator (DWBG) in classification layers to enhance generalization ability. More importantly, owing to the use of depth-wise convolution, the DDUNet is a lightweight network that can achieve 95.3% accuracy on the SWINySEG dataset with only 0.33M parameters, and achieve superior performance over three different configurations of the SWINySEg dataset in both accuracy and efficiency.

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

TL;DR

DDUNet addresses cloud segmentation challenges by combining a U-Net backbone with Dynamic Multi-scale Convolution (DMSC) for adaptive multi-scale feature fusion and Dynamic Weights & Bias Generator (DWBG) for input-specific classifier parameters, all implemented with depthwise convolutions for efficiency. The approach delivers a lightweight model that achieves accuracy with only M parameters on SWINySEG, outperforming several baselines in both accuracy and latency. Across day-time and night-time configurations, the method demonstrates robust performance and strong generalization, making it suitable for real-time deployment on resource-constrained devices. The work highlights how dynamic, parameter-efficient modules can significantly boost cloud segmentation performance without sacrificing speed.

Abstract

Cloud segmentation amounts to separating cloud pixels from non-cloud pixels in an image. Current deep learning methods for cloud segmentation suffer from three issues. (a) Constrain on their receptive field due to the fixed size of the convolution kernel. (b) Lack of robustness towards different scenarios. (c) Requirement of a large number of parameters and limitations for real-time implementation. To address these issues, we propose a Dual Dynamic U-Net (DDUNet) for supervised cloud segmentation. The DDUNet adheres to a U-Net architecture and integrates two crucial modules: the dynamic multi-scale convolution (DMSC), improving merging features under different reception fields, and the dynamic weights and bias generator (DWBG) in classification layers to enhance generalization ability. More importantly, owing to the use of depth-wise convolution, the DDUNet is a lightweight network that can achieve 95.3% accuracy on the SWINySEG dataset with only 0.33M parameters, and achieve superior performance over three different configurations of the SWINySEg dataset in both accuracy and efficiency.
Paper Structure (13 sections, 9 equations, 3 figures, 2 tables)

This paper contains 13 sections, 9 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Architecture
  4. Basic Building Blocks
  5. Dynamic Multi-scale Conv2D (DMSC)
  6. Dynamic Wights & Bias Generator (DWBG)
  7. Loss function
  8. Experiments & Results
  9. Experiments Setting
  10. Qualitative Analysis
  11. Quantitative Analysis
  12. Ablation Study
  13. Conclusion and Discussion

Figures (3)

  • Figure 1: Overall Pipeline, DMSC, and DWBG.
  • Figure 2: Basic building blocks used in DDUNet. (a) DWConv block with $3\times3$ filters; (b) Conv block with $1\times1$ or $3\times3$ filters; (c) Inverted Residual sandler2018mobilenetv2 without expand ratio.
  • Figure 3: Results of cloud segmentation for day-time (1-6 columns) and night-time (7-12 columns).