A complex scene pavement crack semantic segmentation method based on dual-stream framework

Pavement crack detection in complex scenes presents a significant challenge in achieving precise pixel-level accuracy due to the presence of intricate backgrounds, such as water stains, shadows, and repair grouting, which closely resemble cracks in terms of morphology and colour. Consequently, existing approaches often struggle to differentiate these elements, resulting in misclassification and decreased accuracy in corresponding areas. This paper proposes a novel approach, called the Dual-Stream U-Network (DSU-Net), specifically designed for complex scene crack detection. The DSU-Net model employs a dual-stream framework, enabling improved focus on challenging regions and effectively addressing the issue of low recognition accuracy. To construct a comprehensive dataset, photogrammetry techniques are employed to acquire images of cracks, resulting in a dataset comprising 5000 images. Experimental evaluations demonstrate that the trained DSU-Net model achieves a mean F1 score of 74.17% on a test set, surpassing the performance of the state-of-the-art methods which achieved a score of 68.09%. These findings validate the effectiveness of the proposed algorithm in accurately distinguishing water stains, shadows, repair grouting, and cracks. Additionally, the algorithm successfully performs fine-grained extraction of cracks. The promising results obtained in this study indicate the potential practical applicability of the proposed approach in engineering scenarios.