目的 针对物流包装箱几何形变检测中表观干扰易影响结构状态判断的问题,提出一种基于RGB-D视觉的几何形变检测方法。方法 构建由单类箱体候选框定位前端和RGB-D几何验证后端共同组成的检测流程,前端仅根据RGB图像获取箱体候选区域,后端依据深度几何信息完成形变真实性判定。针对待测区域异常深度参与参考深度估计时易引起参考基准偏移的问题,提出自适应环形参考面方法（Adaptive Doughnut Reference Plane,ADRP）,通过外环参考区与中心测量区的空间分离,结合中心-外环深度一致性判断,自适应选择参考深度并构建局部参考平面。结果 在冻结测试集上,系统于低误报运行点取得 P(Precision)=1.000、R(Recall)=0.750、FAR值=0.000、F1=0.857的端到端结果。对比实验与候选框扰动测试结果表明,在人工定锚条件下,ADRP相比固定深度过滤（R=0.036、F1=0.069）和全局 RANSAC（R=0.714、F1=0.833）取得更高的Recall和F1并保持FAR值=0.000。候选框扰动测试进一步说明系统在小范围定位偏移下仍具有一定容忍性。结论 在当前静态俯视RGB-D采集条件和冻结测试集上,所提方法表现出较好的低误报检测能力,可为物流包装箱几何形变自动检测提供RGB-D视觉实现参考。
Abstract
To address the influence of appearance interference on structural-state judgment in geometric deformation detection of logistics packaging boxes, the work aims to propose an RGB-D vision-based detection method. A detection pipeline consisting of a single-class box candidate localization front end and an RGB-D geometric verification back end was constructed. The front end obtained candidate box regions from RGB images only, while the back end determined the authenticity of deformation based on depth geometric information. To avoid reference baseline deviation caused by abnormal depth values in the target measurement region, an Adaptive Doughnut Reference Plane (ADRP) method was proposed. By spatially separating the outer-ring reference region from the central measurement region and introducing a center-ring depth consistency judgment, ADRP adaptively selected the reference depth and constructed a local reference plane. On the frozen test set, the system achieved end-to-end performance of Precision=1.000, Recall=0.750, FAR=0.000, and F1=0.857 at a low-false-alarm operating point. Comparative experiments and candidate-box perturbation tests showed that, under manually anchored regions of interest, ADRP achieved higher Recall and F1 than fixed-depth filtering (Recall=0.036, F1=0.069) and global RANSAC (Recall=0.714, F1=0.833), while maintaining FAR=0.000. The perturbation test further indicated that the system showed tolerance to small localization offsets. Under the current static top-view RGB-D acquisition condition and frozen test set, the proposed method demonstrates a low-false-alarm detection capability and provides an RGB-D vision reference for automatic geometric deformation detection of logistics packaging boxes.
关键词
物流包装箱 /
RGB-D视觉 /
几何形变检测 /
候选区域定位 /
几何验证 /
自适应环形参考面
Key words
logistics packaging box /
Red-Green-Blue-Depth (RGB-D) vision /
geometric deformation detection /
candidate region localization /
geometric verification /
Adaptive Doughnut Reference Plane (ADRP)
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] MA Y, YIN J, HUANG F, et al.Surface Defect Inspection of Industrial Products with Object Detection Deep Networks: A Systematic Review[J]. Artificial Intelligence Review, 2024, 57: 333.
[2] WAHYUDI V, ZIEGLER C C, FRIESS M, et al.A Computer Vision System for Recognition and Defect Detection for Reusable Containers[J]. Machine Vision and Applications, 2025, 36: 39.
[3] WU Y, LU Y.An Intelligent Machine Vision System for Detecting Surface Defects on Packing Boxes Based on Support Vector Machine[J]. Measurement and Control, 2019, 52(7-8): 1102-1110.
[4] YANG M, LI D, LUO P, et al.YOLOv5-S-G-B: A Lightweight Intelligent Detection Model for Cardboard Surface Defects[J]. Signal, Image and Video Processing, 2024, 18: 6997-7011.
[5] AMERI R, HSU C C, SHAMSHIRBAND S, et al.A Systematic Review of Deep Learning Approaches for Surface Defect Detection in Industrial Applications[J]. Engineering Applications of Artificial Intelligence, 2024, 130: 107717.
[6] 宗国浩, 张明琰, 王锐, 等. 卷烟包装外观缺陷数据集构建及深度学习检测技术研究[J]. 包装工程, 2024, 45(5): 135-143.
ZONG G H, ZHANG M Y, WANG R, et al.Construction of Appearance Defect Dataset for Cigarette Package and Research on Deep Learning Detection Technology[J]. Packaging Engineering, 2024, 45(5): 135-143.
[7] 叶宇星, 孙志锋, 马风力, 等. 基于改进YOLOv5s的腌制蔬菜真空包装缺陷检测[J]. 包装工程, 2023, 44(9): 45-53.
YE Y X, SUN Z F, MA F L, et al.Vacuum Packaging Defect Detection of Pickled Vegetables Based on Improved YOLOv5s[J]. Packaging Engineering, 2023, 44(9): 45-53.
[8] CHEN A, LI X, HE T, et al.Advancing in RGB-D Salient Object Detection: A Survey[J]. Applied Sciences, 2024, 14(17): 8078.
[9] HE X, MATSUMARU T.Estimation of Flat Object Deformation Using RGB-D Sensor for Robot Reproduction[J]. Sensors, 2021, 21(1): 105.
[10] DU J, TAO C Y, CAO X M, et al.3D Vision-Based Anomaly Detection in Manufacturing: A Survey[J]. Frontiers of Engineering Management, 2025, 12(2): 343-360.
[11] CAVALIERE G, LANZA O, BORGIANNI Y, et al.Deep Learning-Supported Machine Vision-Based Hybrid System Combining Inhomogeneous 2D and 3D Data for the Identification of Surface Defects[J]. Production & Manufacturing Research, 2024, 12(1): 2378199.
[12] LI Y, LI W, DARWISH W, et al.Improving Plane Fitting Accuracy with Rigorous Error Models of Structured Light-Based RGB-D Sensors[J]. Remote Sensing, 2020, 12(2): 320.
[13] KHOSHELHAM K, OUDE ELBERINK S J. Accuracy and Resolution of Kinect Depth Data for Indoor Mapping Applications[J]. Sensors, 2012, 12(2): 1437-1454.
[14] SAPKOTA R, FLORES-CALERO M, QURESHI R, et al.YOLO Advances to Its Genesis: A Decadal and Comprehensive Review of the You Only Look Once (YOLO) Series[J]. Artificial Intelligence Review, 2025, 58: 274.
[15] FISCHLER M A, BOLLES R C.Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography[J]. Communications of the ACM, 1981, 24(6): 381-395.
[16] ROUSSEEUW P J, CROUX C.Alternatives to the Median Absolute Deviation[J]. Journal of the American Statistical Association, 1993, 88(424): 1273-1283.
[17] CHANLLON. Corrugated Cardboard Boxes Dataset[EB/OL].[2026-06-09]. https://github.com/chanllon/corrugated-cardboard-boxes-dataset/tree/master.
基金
湖州市科学技术局科技计划项目（2025GY013）
PDF(1572 KB)
渝公网安备 50010702501716号 渝ICP备15012534号-2
