Abstract
Reliable visual perception is a prerequisite for autonomous cable duct inspection, particularly for recognizing pipe-joint dislocations and obstruction-related hazards. Images acquired inside cable ducts are often affected by restricted viewpoints, uneven illumination, wall-texture interference, partial occlusion, and substantial variations in target geometry and scale. These factors increase the likelihood of missed targets, false alarms, and inaccurate bounding boxes. This study develops DPR-YOLOv9 from the YOLOv9c detector, where DPR represents deformable-strip feature extraction, position-aware attention, and regression optimization. In the backbone, a Deformable Strip Convolution Network (DSCN) adjusts its sampling pattern to better describe elongated boundaries, displaced joints, and irregular obstacle contours. CoordAttention is introduced into the multi-scale fusion path to retain directional coordinate cues and emphasize spatially relevant features. In addition, Inner-IoU modifies the regression constraint through auxiliary boxes, providing more effective optimization for small or partially occluded targets. Across three independent runs, DPR-YOLOv9 achieved mean Precision, Recall, mAP@0.5, and mAP@0.5:0.95 values of 0.944, 0.933, 0.940, and 0.751, respectively, while maintaining an inference speed of 67.85 FPS. The results indicate that the proposed detector improves both recognition reliability and localization quality for robotic cable duct inspection.