Mamba-enhanced detection transformer with selective state space models for multi-category defect detection in transmission lines

Automated defect detection in transmission lines is critical for power grid safety, yet existing detectors face three key limitations: limited global context in CNNs, quadratic computational complexity $$\mathcal {O}(N^2)$$ in Transformers, and inadequate multi-scale fusion for geometrically diverse defects under long-tailed category distributions. This paper proposes MambaTL-Det, a novel detection framework built upon RT-DETR with four complementary innovations: (1) a MobileMamba backbone leveraging selective state space models to achieve linear $$\mathcal {O}(N)$$ complexity while preserving global modelling capability; (2) a Dynamic Aspect Perception Block (DAP-Block) for aspect-ratio-aware feature enhancement via dual-branch spatial and channel attention; (3) a Strategic Collaborative Optimization Network (SCO-Net) for discriminative cross-scale feature fusion; and (4) a Hierarchical Category-Balanced (HCB) loss that jointly addresses category imbalance, instance difficulty, and geometric diversity. We also construct and release TL-Defect5K, a large-scale dataset comprising 6,908 images with over 25,500 annotations across five defect categories. Extensive experiments demonstrate that MambaTL-Det achieves 87.6% ± 0.3 mAP@0.5 and 62.4% ± 0.2 mAP@0.5:0.95 at 72.5 FPS, consistently outperforming 18 state-of-the-art baselines across four detection paradigms with statistical significance ( $$p<0.05$$ , Wilcoxon signed-rank test over five independent runs). The dataset and code are available at https://github.com/Dali8710/TL-Defect5K .