CMS-RTDETR: multi-scale enhanced real-time detection for underground pipeline defect

Abstract To address the challenges of complex background interference, high miss rates for small object, and insufficient computational efficiency in underground pipeline defect detection, this paper proposes an efficient detection model, CMS-RTDETR, based on improvements to RT-DETR. First, a novel backbone network, cross stage partial-multi-scale edge enhancement, is designed, utilizing multi-scale adaptive pooling and an edge enhancement module to strengthen detail feature extraction, significantly improving detection accuracy for small object and defects with complex textures. Second, an multi-head self-attention module is introduced into the feature interaction layer, integrating multi-scale mechanisms and dynamic attention allocation to effectively enhance contextual information fusion and defect localization robustness in complex backgrounds. Finally, Shape-IoU is employed to replace the traditional regression loss function, optimizing the localization process by incorporating bounding box shape and scale information, thereby reducing regression errors for irregular defects. Experiments show CMS-RTDETR achieves an mAP of 90.4% on a self-built dataset and 82.7% in Sewer-ML generalization tests, improving over RT-DETR-R18 by 2.2% and 2.4%, respectively. Small object detection APs improve by 4.6% and 4.3%. With 20.3% fewer parameters and 56.7 FPS, CMS-RTDETR outperforms mainstream algorithms, balancing accuracy and efficiency for intelligent pipeline inspection.