Open Access

ARTICLE

An Adaptive Multi-Scale Dilated Convolution Network for Real-Time Road Black Ice Detection

Sun-Kyoung Kang¹, Yeonwoo Lee^2,*

1 Department of Computer Software Engineering, Wonkwang University, Jeonbuk, Republic of Korea
2 Department of Artificial Intelligence Engineering, Mokpo National University, Chonnam, Republic of Korea

* Corresponding Author: Yeonwoo Lee. Email:

Computers, Materials & Continua 2026, 88(2), 73 https://doi.org/10.32604/cmc.2026.081553

Received 04 March 2026; Accepted 13 May 2026; Issue published 15 June 2026

Abstract

Black ice formation on road surfaces presents a serious hazard due to its low visibility and high slipperiness, underscoring the critical need for timely and accurate detection in intelligent transportation systems. In this paper, we propose AdaMsDCNet, an adaptive multi-scale dilated convolution network designed for real-time black-ice semantic segmentation on resource-constrained edge platforms, applying a Convolutional Neural Network (CNN) with an adaptive Multi-Scale Dilated Convolution (MsDC) feature fusion encoder-decoder architecture. The key concept of AdaMsDCNet is to employ an encoder-decoder architecture with parallel multi-scale dilated convolutional paths that adjust dilation rates at different encoder depths using a systematic 4→2→1 progression, optimally capturing a wide range of receptive fields while mitigating checkerboard artifacts. The encoder dynamically fuses features from multiple dilation rates at each stage, enhancing segmentation accuracy. Simultaneously, the decoder uses transposed convolutions and skip connections to preserve fine spatial details. Experimental validation on a proprietary thermal infrared dataset of 1156 annotated images show that AdaMsDCNet_9 achieves 96.47% mIoU, 95.48% Black-Ice IoU, 97.55% Precision, 97.82% Recall, and 97.69% F1-Score, outperforming U-Net (+26.78 pp mIoU, +29.88 pp Recall), DeepLabv3+ (+2.82 pp mIoU), and LinkNet (+1.08 pp mIoU) while requiring only 1.86M parameters and maintaining real-time inference speeds of 3.94~5.63 FPS on the NVIDIA Jetson Nano embedded GPU. Ablation studies confirm the benefits of adaptive dilation, parallel feature fusion, and controlled channel growth for the accuracy–efficiency trade-off. Limitations including dataset generalization to uncontrolled outdoor conditions and the evaluation of imbalance-aware loss functions are identified as directions for future work.

---

Keywords

---