A lightweight feature-guided rotation-adaptive oriented detection network for substrate localization in inkjet printing

Random feeding in industrial inkjet printing brings printable substrates into the visual field at variable positions, scales, and in-plane rotations, often accompanied by specular reflection, weak edges, local occlusion, and slight perspective distortion. Before printing, the visual front end must estimate substrate category, center position, size, and angle. This study formulates substrate localization as a lightweight oriented object detection task and proposes FGRA-YOLO, a feature-guided rotation-adaptive detector based on YOLOv11n-OBB. FGRA-YOLO reconstructs local feature responses and rotated directional features through FGRAFusion, which contains a PFGA branch and the proposed SRARConv branch, and combines lightweight feature mapping, a directional geometric enhancement head, KFIoU rotated-box regression, and a double-angle cosine loss (DCL) for explicit orientation supervision. On the completed 2682-image industrial inkjet-printing substrate dataset, FGRA-YOLO achieves mean P, R, mAP50, mAP75, and mAP50-95 values of 90.60%, 92.07%, 94.09%, 94.09%, and 93.80%, respectively. Relative to the single-run YOLOv11n-OBB baseline under the same data split and training protocol, the observed mean mAP50 and mAP50-95 values are higher by 4.35 and 4.41 percentage points, while the parameter count and GFLOPs are lower by 53.1% and 35.4%, respectively. For the printable front and back classes, five FGRA-YOLO seeds obtain detector IoU of 0.9762 ± 0.0026 and center, angle, and maximum-edge errors of 2.515 ± 0.284 px, 0.413 ± 0.070 ∘ , and 6.196 ± 0.643 px, respectively; an identical SAM2 refinement protocol changes these values to 0.9861 ± 0.0000, 1.436 ± 0.004 px, 0.188 ± 0.000 ∘ , and 3.321 ± 0.007 px. The detector contains 1.20 M parameters, requires 10.51 GFLOPs, and runs at 163.02 FPS under the reported test conditions; SAM2 is a separate downstream stage and is excluded from this speed benchmark.