A lightweight feature attention fusion network for pavement crack segmentation

Abstract The occurrence of pavement cracks poses a significant potential threat to road safety, thus the rapid and accurate acquisition of pavement crack information is of paramount importance. Deep learning methods have the capability to offer precise and automated crack detection solutions based on crack images. However, the slow detection speed and huge model size in high‐accuracy models are still the main challenges required to be addressed. Therefore, this research presents a lightweight feature attention fusion network for pavement crack segmentation. This structure employs FasterNet as the backbone network, ensuring performance while reducing model inference time and memory overhead. Additionally, the receptive field block is incorporated to simulate human visual perception, enhancing the network's feature extraction capability. Ultimately, our approach employs the feature fusion module (FFM) to effectively combine decoder outputs with encoder's low‐level features using weight vectors. Experimental results on public crack datasets, namely, CFD, CRACK500, and DeepCrack, demonstrate that compared to other semantic segmentation algorithms, the proposed method achieves both accurate and comprehensive pavement crack extraction while ensuring speed.