摘要
针对传统集装箱表面缺陷检测方法中误报率高和泛化能力不足的问题,本文提出一种基于加权框融合(WBF)与视觉大语言模型(VLLM)审核的两阶段协同检测框架WBF-VLLM。在第一阶段并行引入YOLOv8x与Cascade RCNN模型构建异构视觉检测框架,并采用加权框融合技术对多模型检测结果进行融合;在第二阶段引入视觉大语言模型Gemini 2.5 Pro作为专家审核模块,结合全局与局部图像信息实现候选缺陷进行多模态推理与类别校正。实验结果表明,相较于基准模型,mAP@0.5与mAP@0.5:0.95分别提高45.9%和58.9%,证明所提方法的有效性。
To mitigate the high false positive rate and limited generalization of existing container surface defect detection methods,this paper proposes a two-stage collaborative framework,termed WBF-VLLM,which integrates Weighted Boxes Fusion(WBF)with a Vision Large Language Model(VLLM)-based review.In the first stage,YOLOv8x and Cascade R-CNN are deployed in parallel to form a heterogeneous detection framework,and their outputs are fused via WBF.In the second stage,Gemini 2.5 Pro is introduced as an expert review module to perform multimodal reasoning and category refinement on candidate defects using global and local image information.Experimental results show that the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 45.9%and 58.9%,respectively,over baseline models,which verifies the effectiveness of the proposed method.
作者
杨浩然
王浩然
杨善良
Yang Haoran;Wang Haoran;Yang Shanliang(School of Computer Science and Technology,Shandong University of Technology,Zibo,Shandong 255049,China)
出处
《计算机时代》
2026年第3期49-54,60,共7页
Computer Era
