GTDI-YOLO: An Advanced Single Stage Detector Utilizing Global Topology Decoupling for Dense Object Detection

Abstract

Precisely locating and counting highly dense workpieces on industrial assembly lines remains a critical challenge due to hyper-dense stacking distributions and intense surface specular reflections. To overcome these limitations, we propose GTDI-YOLO, an advanced single-stage detector that reconstructs multi-scale prediction nodes via a plug-and-play global topology decoupled injection head framework. Within this framework, the Multi-Scale Geometric Baseline Alignment module first maps multi-level representations onto a synchronized coordinate frame to preserve micro-positioning sub-pixel clues and fragile reflective textures. Subsequently, the Periodic Topology Cross Reconstruction module shuffles latent representations to isolate feature blending while extracting periodic layout priors and depressing glare noises. Finally, the Gated Multi-Scale Contextual Injection module adaptively routes spatial-gated global context back into scale-specific localized nodes to force sharp activation peaks at target center-points while completely suppressing fake boundary responses. Extensive evaluations on the dense industrial workpiece dataset demonstrate that GTDI-YOLO significantly outperforms state-of-the-art baseline methods in localization accuracy, offering a highly reliable intelligent solution for dense industrial inspection.