**5. Conclusions**

A real-time concrete surface crack detection method based on the improved YOLOv4 is proposed. The improved model for concrete crack detection adopts the symmetry concept in the extraction of backbone and the design of neck and head. It is described in detail in Section 2. A smartphone is used to collect 2000 raw 3024 × 3024 pixel images from the surfaces of concrete buildings. To reduce the computation of the training process, the collected images are cropped to 256 × 256 pixels. Sets of 7000, 1000, and 2000 images are used for training, validation, and testing, respectively. The improved YOLOv4 model achieved an mAP of 94.09%, which is 98.52% of the original YOLOv4 model. The crack detection performance decreased slightly, but the parameters and calculation amount of the model are reduced by 87.43% and 99.00%, respectively. Compared with the results of the high-performance network models in object detection (such as YOLOv4, YOLOv5m, SSD, and CenterNet), it can be concluded that the improved model has almost no loss in mAP, but the model size and calculation amount are greatly reduced. In addition, compared with the detection results of the lightweight network models (such as YOLOv4-tiny and MobileNet-SSD), the model sizes are close, but the calculation amount FLOPs are reduced, and the detection performance mAP is higher. When the improved model was deployed to the Jetson Xavier NX embedded platform for testing, it achieved an mAP of 94.06% with 44 FPS. The size, accuracy, and processing speed of the model can meet the requirements of accurate real-time object detection, which can provide support for the development of mobile monitoring system. As a result, it can achieve real-time automatic vision-based crack detection on concrete surface without other equipment.

Although the improved YOLOv4 model shows good performance, there is still a long way to go before it is suitable for engineering applications. First, in the implementation of the improved method, there are many artificially adjusted hyperparameters derived from the training and verification set. Many experiments need to be conducted to explore the influence of these hyperparameters on the performance of the model. Second, a real-time mobile crack detection system (including APPs and a website) should be developed to monitor the concrete surface cracks for timely repair and protection. Lastly, we will collect more types of defect images to expand the database, such that the proposed method has greater accuracy and robustness.

**Author Contributions:** Conceptualization, G.Y. and Y.S.; methodology, Y.S. and X.L.; software and formal analysis, X.L.; writing—original draft preparation, Y.S.; review and editing, M.W. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the Fundamental Research Funds for the Central Universities (no. 2020CDJQY-A067), the National Key Research and Development Project (no. 2019YFD1101005) and the National Natural Science Foundation of China (no. 51608074).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Please contact the corresponding author.

**Conflicts of Interest:** The authors declare no conflict of interest.
