Research Article
A Systematic Evaluation of Deep Learning Architectures and Training Strategies for Multi-Class Raindrop Segmentation
@INPROCEEDINGS{10.4108/eai.18-12-2025.2365273, author={Kento Ichihara and Lin Meng}, title={A Systematic Evaluation of Deep Learning Architectures and Training Strategies for Multi-Class Raindrop Segmentation}, proceedings={Proceedings of the 13th International Conference on Identification, Information and Knowledge in the Internet of Things, IIKI 2025, 18-21 December 2025, Chengdu, China}, publisher={EAI}, proceedings_a={IIKI}, year={2026}, month={6}, keywords={Computer Vision Semantic Segmentation Deep Learning Raindrop Detection Ablation Study Vision Transformer}, doi={10.4108/eai.18-12-2025.2365273} }- Kento Ichihara
Lin Meng
Year: 2026
A Systematic Evaluation of Deep Learning Architectures and Training Strategies for Multi-Class Raindrop Segmentation
IIKI
EAI
DOI: 10.4108/eai.18-12-2025.2365273
Abstract
Raindrops adhering to camera lenses degrade image quality and impair downstream computer vision tasks in applications such as autonomous driving and surveillance. Most rain detection methods treat all rain as a single class. We introduce multi-class raindrop semantic segmentation, defining four morphological classes: large drops, small drops, blurred drops, and streaks. We benchmark segmentation architectures, encoder backbones including CNNs and Vision Transformers, loss functions, and optimizers on a custom-annotated dataset using 5-fold cross-validation. A U-Net with a Mix Transformer (MiT-B4) encoder, trained with combined Cross-Entropy and Dice loss, AdamW, and Cosine Annealing, achieves a mean IoU of 0.613 and serves as a baseline for this task.
