堆石混凝土(RFC)坝层面保留一定露石能够显著增强层面齿颌与抗剪能力,是施工质量管控的重要环节。然而,如何实现层面露石的准确高效识别与合理评价尚待进一步研究。基于此,提出YOLO-SAM的融合算法框架以增强识别精度与效率,引入基于密度的带有噪声的空间聚类算法DBSCAN以增强层面评价结果的可靠性。研究结果表明:①对于现场采集的24张复杂背景干扰下的层面照片,YOLOv26s-DET+SAM2展现出了较高的识别精度,与人工标注的结果平均误差仅为4.36%,优于YOLOv11s-SEG和YOLOv26s-SEG的YOLO分割模型;②根据规范定义与要求,DBSCAN给出了堆石外露区的判别标准,降低了照片采集范围对评价结果的影响;③不同工程9个层面的分析结果显示,基于YOLO-SAM与DBSCAN算法检测评价方法实现了对堆石外露区的稳定准确判别,有效提升了检测评价结果的可靠性与稳定性。研究成果可为RFC坝现场层面露石的快速检测与质量管控提供技术支撑。

[Objective] Accurate and efficient identification of exposed stones on rock-filled concrete (RFC) dam surfaces is critical for ensuring interlayer shear resistance and construction quality, yet existing methods lack the necessary precision, automation, and reliability under complex onsite conditions. To fill this gap, this study aims to develop a fused algorithm framework that combines YOLO-based detection with the Segment Anything Model (SAM) to significantly improve identification accuracy and efficiency, and to integrate the Density Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to establish a robust, standard compliant criterion for evaluating exposed stone zones. The overarching goal is to provide an automated, repeatable evaluation pipeline that overcomes the limitations of direct segmentation and reduces the influence of photograph collection scope on layer assessment. [Methods] A dataset of 24 highresolution layer photographs was acquired from actual RFC dam construction sites, featuring complex backgrounds including dust, water stains, and uneven lighting. All images were manually annotated at pixel level to generate groundtruth masks of exposed stones. The proposed framework first employs a YOLO variant (v26s-D) as a detector to identify individual stones and produce bounding boxes. Subsequently, SAM2—the latest Segment Anything Model—is applied to each detected region to generate finegrained segmentation masks. The spatial distribution of segmented stone pixels is then analyzed using the DBSCAN clustering algorithm. Based on definitions from relevant construction standards, a clustering criterion is established that groups contiguous exposed stone pixels into meaningful “exposed rock zones” while automatically discarding sparse, isolated noise. The effective exposed stone area ratio is then computed for each layer. For comparison, two YOLObased instance segmentation models (v11s-S and v26s-S) are trained and evaluated on the same data. To assess practical reliability, the complete YOLOSAM2+DBSCAN workflow is further applied to nine independent layers taken from different RFC dam projects. The impact of photograph collection scope on evaluation results is systematically analyzed, and the consistency with expert visual inspections is examined. [Results] On the 24 test photographs, the YOLO(v26s-D)-SAM2 fusion achieved a mean pixel-wise error of only 4.36% compared with manual annotations. This performance clearly surpassed the YOLO segmentation models v11s-S and v26s-S, which exhibited notably higher errors under the same complex backgrounds, demonstrating the advantage of decoupling detection from fine segmentation. With the DBSCAN-based exposed rock zone criterion, the influence of photograph collection scope on the evaluated exposed stone area ratio was substantially reduced; repeated measurements under intentionally varied shooting conditions yielded highly consistent zone boundaries and area percentages, whereas direct full-image segmentation led to obvious fluctuations. In the nine-layer cross-project analysis, the proposed method consistently produced stable and accurate identifications of exposed rock zones. The layer-wise exposed stone ratios closely matched the judgments of experienced site engineers, with discrepancies remaining well within the acceptable tolerance for construction quality control. The method also showed strong repeatability, with a low coefficient of variation across multiple runs, confirming its robustness against real-world disturbances.[Conclusions] The YOLOSAM and DBSCAN fused framework can accurately and efficiently identify exposed stones on RFC dam surfaces, solving the key problem of unreliable segmentation in cluttered construction environments. The 4.36% mean error represents a substantial improvement over conventional YOLO segmentation models, validating the practical value of the detection segmentation cascade. The DBSCAN based postprocessing, explicitly anchored in construction standards, transforms the raw segmentation output into a physically meaningful exposed rock zone evaluation that is largely insensitive to photograph collection scope, thereby greatly enhancing the objectivity and stability of layer assessment. The innovation of this work lies in the synergistic combination of realtime detection, SAM powered fine segmentation,and densitybased unsupervised clustering to achieve both pixellevel accuracy and semanticlevel reliability.