Given the problems of low accuracy of railway container number recognition in complex scenarios and high computational complexity of conventional deep learning models, improved DB and SVTR models for intelligent recognition of railway container numbers based on on-site business rules were proposed. Firstly, in the custom dataset construction stage, data augmentation technologies were adopted to generate a diversified container number dataset containing damaged, occluded, blurred, and nighttime scenarios to enhance the generalization ability of the model. Then, the HAFF module was added after the FPN in the DB detection model to enhance the feature expression ability of key regions, and cross-scale self-attention was combined to improve the text detection ability in complex backgrounds. Secondly, the standard cosine annealing method and Warmup mechanism were used as learning rate update strategies to improve the SVTR recognition model, and the recognition results were verified and corrected through container number encoding rules. Finally, comparative experiments on custom datasets show that the container number detection precision of the improved model reaches 99.0%, and the container number recognition accuracy reaches 95.9%. Meanwhile, the improved SVTR model performs better than the basic SVTR model, CRNN model, and Rosetta model in terms of recognition accuracy and average time consumption.
早期的集装箱箱号自动识别采用条形码技术,因存储量有限而被二维码技术替代,通过射频识别(Radio Frequency Identification,RFID)[1]电子标签传输数据实现自动识别,然而RFID系统存在标准不统一、成本较高和后台兼容性等问题,导致推广受限[2]。而通过使用计算机视觉技术实现集装箱箱号识别,由于其成本较低、部署更方便而得到广泛应用。早期通过形态学操作和连通区域分析等方法[3-6]获取箱号特征,再结合光学字符识别(Optical Character Recognition,OCR)技术将字符特征转换为箱号,但这种方法比较依赖图像质量与光照条件。
深度学习技术的快速发展在相当程度上提升了集装箱箱号自动识别的性能与效果。Jaderberg等[7-8]较早地将卷积神经网络用于自然场景文本检测与字符分类,为后续研究奠定理论基础。Ren等[9]提出的快速区域卷积神经网络(Faster Region-Based Convolutional Neural Networks,Faster R-CNN)框架,提高了检测效率。Wang等[10]通过多尺度特征融合策略,将模型的正确识别率提升至90.4%。Zhong等[11-12]通过引入区域候选网络(Region Proposal Network,RPN)来改进Faster R-CNN目标检测模型,进而提高复杂场景下算法的鲁棒性。崔循等[13-14]通过增加空间注意力机制来改进Faster R-CNN,既保证了检测精准度又兼顾检测速度。为满足检测实时性需求,Liu等[15]提出单次多框目标检测(Single Shot MultiBox Detector,SSD)算法。也有文献针对文本检测研究专门神经网络模型,如连接主义文本提议网络(Connectionist Text Proposal Network,CTPN)[16],将文本行拆分为宽度固定的文本片段,再连接成完整文本行,能有效处理长文本;高效且准确的场景文本检测网络(Efficient and Accurate Scene Text Detector,EAST)[17]直接预测文本行的几何形状,无需候选区域生成,推理速度快;渐进式尺度扩展网络(Progressive Scale Expansion Network,PSENet)[18]采用分割的方法,结合渐进尺度扩展算法能精准分离相邻或重叠文本实例;可微分二值化(Differentiable Binarization,DB)文本检测网络[19]通过动态阈值图学习自适应二值化过程,提升对模糊、低对比度文本的检测精度。
ShileiLYU, LIANGYincong, LIZhen,et al. Orchard Monorail Conveyer Location Based on Ultra High Frequency RFID Dual Antennas and Dual Tags[J]. Transactions of the Chinese Society of Agricultural Engineering,2018,34(4):71-79.
NIUHan, WUXizhi. Pinecone Multi-Target Identification and Location Based on Connected Domain of Otsu Algorithm[J]. Jiangsu Agricultural Sciences,2021,49(15):193-198.
[12]
JADERBERGM, VEDALDIA, ZISSERMANA. Deep Features for Text Spotting[EB/OL]. [2025-01-01].
[13]
JADERBERGM, SIMONYANK, VEDALDIA,et al. Reading Text in the Wild with Convolutional Neural Networks[J]. International Journal of Computer Vision,2016,116(1):1-20.
[14]
RENS Q, HEK M, GIRSHICKR,et al. Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2017,39(6):1137-1149.
[15]
WANGB Y, XUJ Q, LIJ B,et al. Scene Text Recognition Algorithm Based on Faster RCNN[EB/OL]. [2025-01-01].
[16]
ZHONGZ Y, JINL W, HUANGS P. DeepText:A New Approach for Text Proposal Generation and Text Detection in Natural Images[EB/OL]. [2025-01-01].
CHENXinyi, WANGXuejiao, HUShi. Improved Faster R-CNN Text Detection Algorithm[J]. Journal of Anqing Normal University (Natural Science Edition),2021,27(3):67-71.
CUIXun, JINGWenbo, YUHongyang,et al. Container Code Number Location Algorithm Based on Improved Faster R-CNN[J]. Journal of Changchun University of Science and Technology,2020,43(1):45-50.
[21]
WANGZ P, WANGY L. FRCA:High-Efficiency Container Number Detection and Recognition Algorithm with Enhanced Attention[EB/OL]. [2025-01-01].
TIANZ, HUANGW L, HET,et al. Detecting Text in Natural Image with Connectionist Text Proposal Network[EB/OL]. [2025-01-01].
[24]
ZHOUX Y, YAOC, WENH,et al. EAST:An Efficient and Accurate Scene Text Detector[EB/OL]. [2025-01-01].
[25]
WANGW H, XIEE Z, LIX,et al. Shape Robust Text Detection with Progressive Scale Expansion Network[EB/OL]. [2025-01-01].
[26]
LIAOM H, WANZ Y, YAOC,et al. Real-Time Scene Text Detection with Differentiable Binarization[J]. Proceedings of the AAAI Conference on Artificial Intelligence,2020,34(7):11474-11481.
LISijia, KONGDeshun, JIANGHuizeng. Research and Application of Container Handling Informatization Based on Measurement and Detection of Reach Stacker[J]. Railway Transport and Economy,2023,45(3):103-110.
TUHao, DIAOYunfeng, CHENGWenming,et al. Real-Time Container Number Recognition Algorithm Based on Deep Learning[J]. Hoisting and Conveying Machinery,2021(13):58-63.
LIWenhui, YANGWentao, ZHANGYichuan. Construction and Application of Smart Rail Container Management System[J]. Railway Transport and Economy,2021,43(8):43-50.
XIAOJianxiong, LIUBing, CHENGGuangyuan,et al. Application of Packing Photos-Based Intelligent Recognition Technology for China Railway Express Based on Artificial Intelligence[J]. Railway Logistics,2025,43(5):48-52.
LEINinghao, LIUHui, PEIZhengkai,et al. Railway Container Positioning and Number Identification System Based on BeiDou Satellite Navigation System[J]. Railway Computer Application,2025,34(12):50-54.
[42]
DUY K, CHENZ N, JIAC Y,et al. SVTR:Scene Text Recognition with a Single Visual Model[EB/OL]. [2025-01-01].