Airport pavement cracks are characterized by complex morphology,disco ntinuous trend and lots of data noise,and the existing algorithm models have not achieved sati sfactory results.In order to improve the crack detection effect,this paper proposes a new deep lear ning model named "FE-Unet".The model adopts the refine residual block,which not only solves th e problem of gradient transmission in multi-layer network,but also refines the features and integrat es the information of channels,and improves the distinguishing degree of features in each stage.In a ddition,the channel attention block (CAB) in the model can better extract discriminant features and enhan ce the consistency of prediction.Finally,focal loss (FL) is used to make the model focus on the fine crac ks that are difficult to classify.In the experiment,the model is trained with 7 778 ai rport pavement crack images and verified on 1 701 images.In comparison with classical fully convolutional network (FCN),Deep Lab v3 and Unet,FE-Unet has better detection performance for cracks,repair and joint than other models.Amo ng them,the precision,recall rate and F1 value of FE-Unet for crack detection reached 80.3 1%,82.72% and 81.49%.