To address issues such as position and shape uncertainty in pavement crack detection, as well as similarity between crack features and pavement background texture, an improved crack image segmentation algorithm based on multibranch feature shared structure network is proposed. To improve the detection accuracy while reducing the redundancy of computational parameters, a lightweight feature extraction network is used to acquire high level features, and the multibranch hopping connection method is employed to improve the information utilization between channels. Each branch combines the global convolution network (GCN) module and the boundary refinement (BR) module to improve crack edge segmentation and classification robustness within the crack region, and it employs the recurrent residual convolution (RRC) module to drive crack feature accumulation. Furthermore, the crack morphological parameters are calculated using the median-axis method to extract the crack skeleton, and relative errors of the crack length and width are 4.73% and 5.21%, respectively. The results of multiple datasets of designed comparison experiments show that the proposed improved algorithm can significantly improve the accuracy and efficiency of pavement crack detection.