目的 针对传统人工检测效率低、成本高且难以规模应用的局限,提出一种基于深度学习的包裹损坏自动识别算法。该算法主要面向物流环节中的分拣中心与转运节点,可在包裹进入自动化分拣线前后,基于实时拍摄的图像,对包装完好性进行实时判断,从而为拦截、返修或重新包装等后续处理提供依据,避免破损包裹继续流转导致货损。方法 研究采用的数据来源包括公开数据集、网络采集及自行采集图像,涵盖完好、污损、弯折和破损4类包裹状态。在迁移学习策略下,对多种主流深度学习模型进行对比分析,发现 GoogLeNet 综合性能最优。因此,以其为基线网络,并在此基础上引入CBAM 注意力机制、LeakyReLU 激活函数、DropBlock 正则化、Bottleneck 残差结构及分类头重构,同时结合标签平滑交叉熵损失、AdamW优化器与 OneCycleLR 学习率调度策略对模型进行优化。结果 改进后的GoogLeNet 在测试集上取得94.36%的分类准确率,Precision、Recall与F1-Score 均较基线模型明显提升,且在模型规模与推理时间方面保持了较好的效率表现。结论 研究结果验证了本文提出的多策略优化的改进GoogLeNet在快递包裹损坏识别任务中的有效性与实用性,可为智能物流分拣与仓储质检等应用场景提供可靠的技术参考。
Abstract
To address the limitations of traditional manual inspection, which is inefficient, costly, and difficult to scale, the work aims to propose a deep learning-based automatic package damage detection algorithm. The algorithm is primarily designed for logistics nodes such as sorting centers and transfer hubs. By analyzing images captured in real time before and after packages enter automated sorting lines, it enables real-time assessment of package integrity and provides a basis for subsequent actions such as interception, repair, or repackaging, thereby preventing further handling and transit of damaged packages and reducing potential losses. Data were collected from public datasets, web sources, and self-captured images, covering four conditions: intact, soiled, bent, and broken. Through a transfer learning approach, several mainstream deep learning models were compared, with GoogLeNet showing the best overall performance and thus selected as the baseline. The model was subsequently enhanced by integrating the CBAM attention mechanism, LeakyReLU activation, DropBlock regularization, a Bottleneck residual structure, and a redesigned classifier head. Further optimizations included label-smoothed cross-entropy loss, the AdamW optimizer, and a OneCycleLR learning rate scheduler. The improved GoogLeNet achieved a classification accuracy of 94.36% on the test set, with marked gains in Precision, Recall, and F1-Score over the baseline, while maintaining favorable efficiency in model size and inference time. These findings confirm the effectiveness and practicality of the proposed multi-strategy optimized GoogLeNet for package damage recognition, offering a reliable technical reference for intelligent logistics sorting and quality inspection in warehouses.
关键词
深度学习 /
快递包裹损坏识别 /
迁移学习 /
注意力机制 /
GoogLeNet
Key words
deep learning /
package damage recognition /
transfer learning /
attention mechanism /
GoogLeNet
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] 伍梓鸿. 基于知识图谱赋能快递包装治理研究[J]. 包装工程, 2023, 44(S1): 230-239.
[2] WU Z H.Research on Empowering Express Packaging Governance Based on Knowledge Map[J]. Packaging Engineering, 2023, 44(S1): 230-239.
[3] 高思伟. 基于机器视觉的包装缺陷检测研究[D]. 太原: 太原科技大学, 2023.
[4] GAO S W.Research on Packaging Defect Detection Based on Machine Vision[D]. Taiyuan: Taiyuan University of Science and Technology, 2023.
[5] 彭小鹏, 黎彩怡, 尚华. 一种可拆卸回收的快递包装箱结构设计研究[J]. 包装工程, 2022, 43(17): 196-202.
[6] PENG X P, LI C Y, SHANG H.Structure Design of a Removable and Recyclable Express Packaging Box[J]. Packaging Engineering, 2022, 43(17): 196-202.
[7] 国家邮政局. 关于2023年四季度邮政业用户申诉情况的通告[EB/OL]. (2024-02-07)[2026-01-09]. https://www.spb.gov.cn/gjyzj/c100015/c100016/202402/4c3199f6285642afae84aae9e237d9ee.shtml.
[8] State Postal Bureau. Notice on the Complaints of Postal Industry Users in the Fourth Quarter of2023[EB/OL]. (2024-02-07) [2026-01-09]. https://www.spb.gov.cn/gjyzj/c100015/ c100016/ 202402/4c3199f6285642afae84aae9e237d9ee.shtml.
[9] 青海省邮政管理局. 2025年11月全省邮政业消费者申诉情况通报[EB/OL]. (2025-12-11)[2026-01-09]. https://qh.spb.gov.cn/qhsyzglj/c100057/c100058/202512/9f4989014a3948e095e92cf133027284.shtml.
[10] Qinghai Postal Administration. Notification on Consumer Complaints in the Provincial Postal Industry for November2025[EB/OL]. (2025-12-11)[2026-01-09]. https://qh.spb.gov.cn/qhsyzglj/c100057/c100058/202512/9f4989014a3948e095e92cf133027284.shtml.
[11] 胡艺川, 陈兴旺, 王姝, 等. 基于双通道残差的工业产品表面缺陷检测方法[J]. 包装工程, 2025, 46(13): 220-232.
[12] HU Y C, CHEN X W, WANG S, et al.Method of Industrial Product Surface Defect Detection Based on Dual-Channel Residual Networks[J]. Packaging Engineering, 2025, 46(13): 220-232.
[13] LENG Y F, LIU J Z.Improved Faster R-CNN for Steel Surface Defect Detection in Industrial Quality Control[J]. Scientific Reports, 2025, 15: 30093.
[14] YANG S, ZHANG M M, YANG Y J, et al.Improving the Positioning Accuracy of Surface Defect Recognition in Power Equipment Using the Improved Faster R-CNN Model[J]. Journal of Physics: Conference Series, 2025, 3077(1): 012013.
[15] 白先浪, 张群利, 辛志强. 融合注意力机制和加权双向特征网络的皮革包装缺陷检测方法[J]. 包装工程, 2025, 46(17): 232-242.
[16] BAI X L, ZHANG Q L, XIN Z Q.Leather Packaging Defect Detection Method Integrating Attention Mechanisms and Weighted Bidirectional Feature Network[J]. Packaging Engineering, 2025, 46(17): 232-242.
[17] 吐尔逊·买买提, 邱建卓, 朱兴林, 等. 基于YOLOv8n的轻量化道路裂缝检测算法[J]. 现代电子技术, 2025, 48(13): 11-19.
[18] TURSUN M, QIU J Z, ZHU X L, et al.Lightweight Road Crack Detection Algorithm Based on YOLOv8n[J]. Modern Electronics Technique, 2025, 48(13): 11-19.
[19] 姜红, 杨棋驭, 张馨艺. 基于DRS-PCA-深度森林架构对牛皮纸袋的分类研究[J]. 包装工程, 2025, 46(17): 265-270.
[20] JIANG H, YANG Q Y, ZHANG X Y.Classification of Kraft Paper Bags Using DRS-PCA Deep Forest Architecture[J]. Packaging Engineering, 2025, 46(17): 265-270.
[21] KIM S, LEE S D.YOLO-Based Damage Detection with StyleGAN3 Data Augmentation for Parcel Information-Recognition System[J]. Computers, Materials & Continua, 2024, 80(1): 195-215.
[22] CHEN Z, DU C F, GUAN Q L, et al.Efficient Parcel Damage Detection viaFaster R-CNN: A Deep Learning Approach forLogistical Parcels’ Automated Inspection[C]// Mobile and Ubiquitous Systems: Computing, Networking and Services. Cham: Springer, 2024.
[23] NAUMANN A, HERTLEIN F, DÖRR L, et al. Parcel3D: Shape Reconstruction from Single RGB Images for Applications in Transportation Logistics[C]// 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Vancouver: IEEE, 2023.
[24] LI Z Q, AI Q, PENG E D, et al.LAP-Net: A Lightweight PCB Defect Detection Network Combined with Attention Mechanisms[J]. Journal of Real-Time Image Processing, 2025, 22(5): 166.
[25] 赵敏, 范英, 高思伟, 等. 结合Swin-Transformer的改进YOLOv5s包装盒缺陷检测算法[J]. 制造业自动化, 2024, 46(12): 34-40.
[26] ZHAO M, FAN Y, GAO S W, et al.Improved YOLOv5s Packaging Defect Detection Algorithm Combined with Swin-Transformer[J]. Manufacturing Automation, 2024, 46(12): 34-40.
[27] ZHUANG F Z, QI Z Y, DUAN K Y, et al.A Comprehensive Survey on Transfer Learning[J]. Proceedings of the IEEE, 2021, 109(1): 43-76.
[28] 赵一璇, 原晓楠, 王戈, 等. 基于GoogLeNet卷积神经网络的智能垃圾分类系统设计[J]. 实验科学与技术, 2025, 23(5): 54-59.
[29] ZHAO Y X, YUAN X N, WANG G, et al.Design of an Intelligent Garbage Classification System Based on the GoogLeNet Convolutional Neural Network[J]. Experimental Science and Technology, 2025, 23(5): 54-59.
[30] PATIL R A, DIXIT V V.Detection and Classification of Mammogram Using ResNet-50[J]. Multimedia Tools and Applications, 2025, 84(31): 38287-38310.
[31] CHENG X C, WU M L, FENG X, et al.ConvNeXt-Driven Dynamic Unified Network with Adaptive Feature Calibration for End-to-End Person Search[J]. Computers, Materials & Continua, 2025, 85(2): 3527-3549.
[32] WU Y.A Lightweight Modulation Recognition Network Based on MobileNetV4[C]// 2025 International Conference on Communications, Computing and Artificial Intelligence. Nanjing: IEEE, 2025.
[33] 林冬梅, 王鑫瑜, 杨玉荷, 等. 基于EfficientNetV2-多注意力机制的脉象识别[J]. 计算机技术与发展, 2025, 35(10): 173-180.
[34] LIN D M, WANG X Y, YANG Y H, et al.Pulse Recognition Based on EfficientNetV2-Multiattention Mechanism[J]. Computer Technology and Development, 2025, 35(10): 173-180.
[35] FU W H, HE Y C, LIU X J, et al.Deep Learning-Enabled Fault Diagnosis with Vision Transformers and Domain-Informed Data Augmentation for Advanced Quality Control in Semiconductor Manufacturing[J]. Applied Soft Computing, 2026, 188: 114480.
[36] M D, B R A P. Multi-Phase Deep Learning Framework with Multiscale Adaptive Swin Transformer and Embedding Attention for Precision Lung Nodule Detection and Classification[J]. Scientific Reports, 2026, 16: 1652.
[37] 赵泮真, 王松峰, 齐飞, 等. 基于迁移学习的MobileViT-CBAM鲜烟叶成熟度识别模型研究[J]. 中国烟草科学, 2025, 46(2): 93-100.
[38] ZHAO P Z, WANG S F, QI F, et al.MobileViT-CBAM Model for Fresh Tobacco Leaf Maturity Recognition Based on Transfer Learning[J]. Chinese Tobacco Science, 2025, 46(2): 93-100.
[39] TANWAR V, SHARMA B, YADAV D P, et al.Hybrid Deep Learning Framework Based on EfficientViT for Classification of Gastrointestinal Diseases[J]. Scientific Reports, 2025, 15: 26982.
[40] 吴聪, 郭志强, 杨杰. 基于改进的注意力机制残差网络穴盘幼苗分类算法研究[J]. 激光与光电子学进展, 2022, 59(22): 81-90.
[41] WU C, GUO Z Q, YANG J.Algorithm for Plug Seedling Classification Based on Improved Attention Mechanism Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(22): 81-90.
[42] ZHOU D Q, HOU Q B, CHEN Y P, et al.Rethinking Bottleneck Structure for Efficient Mobile Network Design[C]// Computer Vision-ECCV 2020. Cham: Springer, 2020.
[43] HUANG G, LI Y, PLEISS G, et al.Snapshot Ensembles: Train 1, Get M for Free[C]// International Conference on Learning Representations. Toulon: ICLR, 2017.
[44] 廖新芝, 孔国希, 林桂潮, 等. 基于CBAM-YOLOv8的温室番茄果实识别[J]. 中国瓜菜, 2025, 38(9): 48-56.
[45] LIAO X Z, KONG G X, LIN G C, et al.Greenhouse Tomato Fruit Recognition Based on CBAM-YOLOv8[J]. China Cucurbits and Vegetables, 2025, 38(9): 48-56.
[46] JI S K, HAO S Y, CUI J C, et al.Identification of Syrup Adulteration in Wolfberry Honey Using CNN-CBAM-SVM Combined with 1H NMR[J]. Food Chemistry, 2025, 493: 145728.
[47] DUAN L J, ZHANG Z C, LIU Z Y, et al.A Weakly-Supervised Oriented Object Detector: Knowledge-Based Dropblock and Unified Regression Network[J]. Neural Networks, 2025, 191: 107830.
[48] GU S, ZHANG G Y, JIA C W, et al.Attention-Based Batch Normalization for Binary Neural Networks[J]. Entropy, 2025, 27(6): 645.
[49] MAO A, MOHRI M, ZHONG Y.Cross-Entropy Loss Functions: Theoretical Analysis and Applications[C]// International Conference on Machine Learning. Honolulu: PMLR, 2023.
PDF(7605 KB)
渝公网安备 50010702501716号 渝ICP备15012534号-2
