Open Access

ARTICLE

DSC-RTDETR: An Improved RTDETR Based Crack Detection on Concrete Surface

Yan Zhou, Hengyang Wu^*

School of Computer and Information Engineering, Shanghai Polytechnic University, Shanghai, 201209, China

* Corresponding Author: Hengyang Wu. Email:

Journal on Artificial Intelligence 2025, 7, 381-396. https://doi.org/10.32604/jai.2025.071674

Received 09 August 2025; Accepted 12 September 2025; Issue published 20 October 2025

Abstract

Crack Detection is crucial for ensuring the safety and durability of buildings. With the advancement of deep learning, crack detection has increasingly adopted convolutional neural network (CNN)-based approaches, achieving remarkable progress. However, current deep learning methods frequently encounter issues such as high computational complexity, inadequate real-time performance, and low accuracy. This paper proposes a novel model to improve the performance of concrete crack detection. Firstly, the You Only Look Once (YOLOv11) backbone replaces the original Real-Time Detection Transformer (RTDETR) backbone, reducing computational complexity and model size. Additionally, the Dynamic Snake Convolution (DSConv) has been introduced to strengthen the model’s ability to extract crack features. Furthermore, integrating Cross-Stage Partial + Parallel Spatial Attention (C2PSA) and BiFormer attention mechanism to enrich the extracted feature information and detection accuracy. Experimental results demonstrate that compared with the RTDETR baseline model, the proposed method achieves a 1.8% improvement in detection accuracy and a 2.1% increase in recall, while reducing computational complexity by 55.3 GFLOPs and parameter count by 14 M. These results highlight the superior performance and efficiency of the proposed model.

---

Keywords

---