Cracks are a typical distress in airport runway pavements. Timely detection is crucial for enhancing pavement performance and preventing their progression into structural damage. To address the current issues of insufficient detection accuracy under low illumination and noise interference, as well as limited ability to identify cracks of different scales in runway crack detection, an improved AER-UNet detection model was proposed. Three aspects of the proposed method were improved based on the UNet model. Firstly, an ASPP module was integrated at the end of the encoder to enhance the capability for multi-scale crack feature extraction. Secondly, an ECA attention mechanism was embedded into skip connections to improve robustness against noise and low-contrast backgrounds. Finally, the standard convolutions of the UNet model were replaced with residual convolution blocks to alleviate the vanishing gradient problem. In addition, a crack parameter quantification algorithm was proposed to address the limitation of existing methods that only provide segmentation results. Comparative experiments and visualization analyses demonstrate that the performance of AER-UNet was superior to that of the original UNet and other comparative models. Its precision, recall, and F₁ score reach 85.7%, 84.5%, and 85.1%, respectively. Furthermore, crack parameter quantification algorithm possesses high accuracy, which can provide data support for the airport runway crack damage assessment and scientific maintenance decision-making.