A single-image desnowing model with joint foreground-background learning
CSTR:
Author:
Affiliation:

(1.College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China; 2.Hytera Science & Technology Co., Ltd., Shenzhen 518057, Guangdong, China)

Clc Number:

TP391.4

Fund Project:

  • Article
  • |
  • Figures
  • |
  • Metrics
  • |
  • Reference
  • |
  • Related
  • |
  • Cited by
  • |
  • Materials
  • |
  • Comments
    Abstract:

    To address the issue that current single-image desnowing models generally only focus on the modeling and removal of foreground snowflakes while neglecting image background information, resulting in low overall restoration accuracy, a single-image snow removal model based on joint foreground-background learning is proposed. Firstly, a dual-scale phantom feature module based on depth wise convolution is designed to enhance the representation ability of snowflakes and background details while effectively mitigating the risk of overfitting; Secondly, a joint foreground-background learning mechanism is proposed to introduce background semantic information into the snowflake modeling process, guiding the model to comprehensively understand the image structure and scene context; Finally, a bilateral long-short-term feature fusion unit is constructed. To promote full fusion and interaction between foreground and background features,which further improves the quality and stability of image restoration. Effectiveness was evaluated on two public benchmark datasets Snow100K and CSD. The results show that the proposed model outperforms existing mainstream models in terms of RPSN, SSSIM, demonstrating stronger robustness, generalization ability, and ability to process complex snow scene images. This model effectively compensates for the background modeling deficiencies in existing models and has high practical value for real-world scenarios.

    Reference
    Related
    Cited by
Get Citation
Related Videos

Article Metrics
  • Abstract:
  • PDF:
  • HTML:
  • Cited by:
History
  • Received:October 12,2024
  • Revised:
  • Adopted:
  • Online: October 11,2026
  • Published:
Article QR Code