A crack detection method based on improved Unet model (A-Unet) is proposed to solve the problems of complex concrete cracks and many interference factors in bridges, tunnels and other environments. Firstly, Unet-based network, how the deep of the encoder affects the training time and detection accuracy of the model is studied. Secondly, in the decoder process, a fusion space and channel attention module is designed to give different weights to the high-resolution shallow features and the deep feature information obtained from the up-sampling to further enhance the crack features. At the same time, the dice loss function is added to evaluate the model to reduce the problem of inaccurate evaluation caused by the large difference between the number of detected objects and the background. The proposed method was evaluated in the test data set, the Precision, MIou and Recall rate reached 94.70%, 86.16% and 91.34% respectively. Also, the detection effect of A-Unet model is significantly better than the other five models. The results show that the accuracy of concrete crack detection by this method is greatly improved, and the model training time is saved, and the detection efficiency is improved.