Abstract: The complex surface image conditions of bridges seriously reduce the crack recognition accuracy of the intelligent identification algorithm. To address this issue, a two-stage bridge crack identification algorithm is proposed. The network structure of MobilenetV3 is adjusted according to the characteristics of bridge surface images, and a classification model for bridge surface images is constructed to achieve automatic filtering of background images, interfering images and low-quality images. A box-grid fusion crack identification model is trained to achieve crack locating, which is used for further crack segmentation and crack measurement. The final image classification accuracy is 97.9%, and the final crack recognition accuracy for grid prediction is 65.6%. The results indicate that the intelligent detection method for bridge cracks proposed in this article can achieve efficient and automated detection of bridge cracks, and the algorithm has strong anti-interference ability.