Hierarchical spatio-temporal adaptive deraining network for real-time autonomous driving applications
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
With the widespread deployment of autonomous driving systems, vehicle perception must remain robust under adverse weather conditions. Traditional raindrop removal algorithms degrade with real-time windshield occlusions and atmospheric streaks, impairing vehicle and pedestrian detection. We propose a lightweight network addressing these challenges through attention mechanisms and adaptive multi-scale feature fusion. A lightweight spatial attention module reduces complexity from O(N2) to O(N) while capturing fine-grained spatial dependencies essential for perception. The adaptive multi-scale extraction architecture handles diverse raindrop patterns with parallel paths, asymmetric convolution, and dilated convolution. A parallel dual-attention color preservation mechanism maintains red, green, blue consistency through independent spatial and channel attention, and a multi-color space loss constrains reconstruction errors in hue, saturation, value and luma, chrominance spaces and preserves white balance. Experiments on Rain100H, Rain100L, and RainDrop demonstrate that hierarchical spatio-temporal adaptive deraining network outperforms existing methods, achieving 42.62 dB peak signal-to-noise ratio on Rain100L and 31.82 dB on RainDrop, sustaining 67.29 frames per second with 43.3 M.
Fields of Study & Subjects
Image Enhancement Techniques
Image and Signal Denoising Methods
Advanced Image Processing Techniques
Publication details
- Journal / source
- Journal of Electronic Imaging
- Volume
- 35
- Issue
- 02
- Pages
- 1–20
- Publication date
- 27 March 2026
- Language
- EN