Enhanced YOLOv10 for Small-Object Rebar Detection in UAV Images on Construction Sites

초록

Unmanned Aerial Vehicles (UAV) images can streamline reinforced-concrete inspections, yet reliable automatic rebar counting is still challenging because rebars often occupy only a few pixels, appear in dense groups, and are frequently obscured or visually blended with background textures, shadows, and occlusions. To address this small-object setting, we develop an improved You Only Look Once (YOLO)v10-based detector that incorporates Omni-Dimensional Dynamic Convolution (ODConv) into the backbone, an Efficient Multi-scale Attention-guided Bidirectional Feature Pyramid Network (EMA-BiFPN) to strengthen multi-scale feature aggregation, a four-scale prediction design that adds an extra high-resolution detection head, and a Minimum Points Distance IoU (MPDIoU) regression loss to stabilize bounding-box learning under stringent IoU criteria. Experiments on a dataset collected from seven construction sites (1,328 images) show that the proposed approach achieves AP50 = 94.72 and AP50:95 = 71.86 at 35.93 FPS, outperforming the YOLOv10 baseline and delivering clearer gains for extremely small rebar instances.

키워드