集装箱重量字符实时视觉识别方法  

Real-time Visual Recognition of Container Weight Characters

作　　者：涂铮 万志伟[2] Tu Zheng;Wan Zhiwei(Wuhan Guide Intelligent Technology Co.,Ltd.;School of Automation,Wuhan University of Technology)

机构地区：[1]武汉港迪智能技术有限公司,武汉430220 [2]武汉理工大学自动化学院

出　　处：《港口装卸》2022年第6期39-43,56,共6页Port Operation

摘　　要：针对集装箱后箱面重量和容积字符信息精确自动识别问题,提出一种集装箱重量字符实时视觉识别方法。运用改进的像素级分割文本检测方法(DBNet)以及改进的二维注意力文本识别方法(SAR),分别完成集装箱重量字符的检测和识别任务,后处理模块融合文本识别结果,得到最终需要获取的重量和容积信息。实验证实,该方法在检测子任务上相较原始DBNet方法推理速度提升27%,F-score提升5.8;识别子任务上,相较原始SAR网络,推理速度提升42%,精度仅降低0.4%。在最终的重量信息检验实验中,仰视视角拍摄下的数据集识别精度达到97.6%,FPS为20.2帧/s,基本达到研究预期,验证了方法的有效性。A real-time visual recognition method for container weight characters is proposed to solve the problem of automatic recognition of container weight and volume characters.The improved pixel level segmentation text detection method(DBNet)and the improved two-dimensional attention text recognition method(SAR)are used to complete the detection and recognition tasks of container weight characters respectively.The post-processing module fuses the text recognition results to obtain the final weight and volume information that needs to be obtained.Experiments show that the reasoning speed of this method is 27%faster than that of the original DBNet method in detecting subtasks,and F-score is 5.8 higher than that of the original DBNet method.In the recognition sub task,compared with the original SAR network,the reasoning speed is improved by 42%,and the accuracy is only reduced by 0.4%.In the final weight information verification experiment,the recognition accuracy of the dataset under the upward viewing angle is 97.6%,and the FPS is 20.2 frames/s,which basically meets the research expectation.And the effectiveness of the method is verified.

关 键 词：智能交通 集装箱重量字符 实时视觉文本识别 文本检测 

分 类 号：U695.22[交通运输工程—港口、海岸及近海工程] TP391.41[交通运输工程—船舶与海洋工程]