Abstract: To address the issues of insufficient sample data and poor image quality in the crack recognition task based on deep learning, a study on crack image enhancement technology using generative adversarial network (GAN) and superresolution (SR) reconstruction is conducted. Firstly, a basic crack database is established using the public crack datasets from tunnel crack and CFD. Secondly, the sample set is expanded by methods such as mirror transformation, brightness adjustment, and random masking, and the model is improved using SA-BAGAN-GP for adversarial generation. The self-attention mechanism is combined to effectively enhance the structural consistency and authenticity of the generated images. Thirdly, to address the problem of blurry details in the generated images, a super-resolution reconstruction algorithm based on Real-ESRGAN is introduced to significantly enhance the crack edges and details. Finally, the data is labeled using labelme to construct a high-quality dataset containing 2400 crack images and corresponding labels, providing reliable data support for the optimization of the crack recognition algorithm.