Abstract
Abstract
Accurate pavement crack segmentation is essential for intelligent road inspection and preventive pavement maintenance. However, existing RGB-based semantic segmentation methods are highly sensitive to illumination variations, while conventional RGB–Depth fusion strategies often fail to effectively exploit complementary appearance and geometric information. To address these limitations, this paper proposes CFUNet, a novel dual-branch RGB–Depth semantic segmentation framework based on a mid-level fusion strategy. The proposed network adopts two independent ResNet50 encoders to extract modality-specific features from RGB images and depth maps. An Enhanced Weighted Fusion (EWF) module is introduced to perform adaptive cross-modal feature aggregation, while a Lightweight Fusion Vision Transformer (LFVit) enhances global semantic representation with reduced computational complexity. Furthermore, a Frequency-aware Upsampling (FAUpSample) module is designed to preserve high-frequency boundary information during decoder reconstruction. Extensive experiments on a public RGB–Depth pavement crack dataset demonstrate that the proposed method consistently outperforms representative CNN-based, Transformer-based, and multi-modal segmentation networks in terms of Precision, Recall, F1-score, and Intersection over Union (IoU). Ablation studies further verify the effectiveness of each proposed module and demonstrate that their collaborative optimization substantially improves segmentation performance. The proposed CFUNet provides an accurate and computationally efficient solution for intelligent pavement crack segmentation and shows considerable potential for deployment in practical road inspection systems.