Comparative Study of Fusion Strategies for Pavement Distress Detection with Intensity and Range Images

With the continuous expansion of China’s highway network, the task of road maintenance has become increasingly demanding. Pavement distress requires early identification, but single-image modalities—such as intensity and range images—are insufficient to comprehensively capture both the spatial texture and depth characteristics of distress, highlighting the need for complementary information. To address this issue, we systematically investigate fusion methods for heterogeneous pavement images, aiming to enhance the accuracy and efficiency of distress identification through image fusion. Focusing on fusion strategies for pavement distress detection, we build up networks to compare different fusion strategies. In early fusion, multiple image-level fusion methods—including pixelwise averaging, target-aware dual adversarial learning (TarDAL), and a dual-branch feature decomposition fusion network (DAF-Net)—were applied to generate fused images for direct detection. In intermediate and late fusion, dual-branch networks capable of simultaneously processing intensity and range images were constructed, integrating several attention-based feature fusion modules such as content-guided attention fusion (CGAFusion), inverted residual mobile block fusion (iRMBFusion), profound semantic fusion module (PSFM), and the proposed synergistic spatial, channel, and pixel attention fusion (SCPFusion) module. The SCPFusion module combines three attention mechanisms—spatial, channel, and pixel attention—to achieve adaptive fusion and enhancement of features through dynamic weighting. The proposed SCPFusion module achieved the highest detection performance under the late fusion, with particularly outstanding performance in transverse crack detection. The results explore fusion strategies and methods for heterogeneous pavement images, emphasizing its potential for practical applications in infrastructure maintenance.