DeepLabV3+_KBSS for lightweight pavement crack segmentation and quantification in complex scenarios

Accurate pavement crack detection and quantitative assessment are essential for road infrastructure maintenance. However, existing segmentation methods suffer from limited robustness under complex illumination, insufficient modeling of elongated crack continuity, and a lack of automated pipelines for engineering-grade evaluation. To address these challenges, this paper proposes DeepLabV3+_KBSS, a lightweight crack segmentation and assessment framework. The model integrates three complementary modules in a position-aware cascade: a KAN Pyramid Enhancement Module for adaptive illumination normalization, a post-ASPP BiFormer module for capturing long-range crack dependencies via bi-level routing attention, and a decoder-embedded SaE module for optimizing edge-semantic feature fusion. Based on the segmentation results, a skeleton-based post-processing pipeline is established to extract geometric parameters—length, width, and area—and compute the Pavement Condition Index (PCI) in accordance with the JTG 5210 − 2018 standard. Experiments are conducted on a mixed dataset comprising 4,029 images (3,000 from the public Crack-BPHDR dataset and 1,029 from field-collected road scenes), with an 8:1:1 train-validation-test split. The proposed method achieves an mIoU of 82.58%, mAP of 90.28%, mRecall of 89.21%, and mDice of 89.88%, with only 3.18 M parameters and an inference speed of 108 FPS, outperforming CNN-based and attention-based methods. These results demonstrate that the proposed method provides an effective, lightweight, and deployable solution for automated pavement distress inspection and maintenance decision-making.