The road crack detection algorithm improved based on YOLOv7

作者：Xin Liang, Maoting Gao · 年份：2024 · DOI：10.1145/3665348.3665405 · 研究领域：Infrastructure Maintenance and Monitoring、Non-Destructive Testing Techniques、Concrete Corrosion and Durability

We propose an algorithm for road crack detection, leveraging enhancements to YOLOv7. This approach aims to overcome challenges related to the limited capacity for extracting features from irregularly shaped road crack targets and the sluggish inference speed caused by the intricate model structure. In this algorithm, YOLOv7 is used as the foundation. Firstly, the Detect module is improved to a Decoupled Head Detect module to augment the detection capacity of the model on feature maps. Then, the Head part is augmented with an ELAN-G module based on GSConv to reduce the model's parameter computation and accelerate the inference speed. Finally, Deformable Conv is employed to modify the original ELAN module into an ELAN-D module and integrate it into the Backbone part to bolster the model's capability in extracting irregular-shaped crack targets. Experimental results on the CRDDC-2020-US dataset demonstrate that the improved algorithm achieves a 10% enhancement in detection accuracy, with the mAP(0.5) increasing from 53.7% to 59.4%. Furthermore, the model's parameter computation is decreased by 12%, while the throughput in terms of the number of images processed per second during inference is boosted by 13%. This algorithm proves to be effective for road crack detection applications.