An Improved RT-DETR-Based Algorithm for Slender Object Detection

Le résumé fourni par la source

Slender object detection plays a crucial role in application scenarios such as industrial automation, intelligent transportation, and UAV inspection. However, due to the unique morphological characteristics and large scale variations inherent to such targets, current detection models still face bottlenecks in balancing accuracy and real-time performance. This study proposes an optimized model based on RT-DETR (Real-time Detection Transformer), dedicated to enhancing the detection efficacy of slender objects. Specifically, by embedding a Cascaded Group Attention (CGA) mechanism — which groups attention heads and establishes cascaded information flow between them to progressively capture local-to-global structural coherence of slender objects—the model's perception capability for both local details and global structures of slender targets is strengthened. Meanwhile, a Hierarchical and Selective Feature Pyramid Network (HS-FPN) architecture is adopted to reconstruct the feature extraction process, enhancing the model's adaptive capability and robustness for multi-scale slender targets. Experiments conducted on datasets covering multiple types of slender objects demonstrate that the proposed method significantly outperforms existing mainstream detection algorithms in key metrics such as precision (67.9%), recall (51.1%), and mean Average Precision (mAP@0.5 of 57.6%), while maintaining excellent real-time performance (38 FPS). This research provides an efficient and accurate technical approach for slender object detection tasks, with its effectiveness and feasibility validated through practical application scenarios. The proposed model achieves a favorable balance between accuracy and speed, making it suitable for real-time applications in industrial, transportation, and biomedical imaging systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.