【作者】 梁艳；

【导师】 姚寒冰；

【作者基本信息】 武汉理工大学 ， 计算机科学与技术， 2020， 硕士

【摘要】 随着港口运输越来越发达,为适应港口集装箱业务的发展需求,集装箱作业自动化的程度也越来越高,正确高效地识别集装箱箱号是自动化作业的基础。随着人工智能在港口领域得到广泛应用,机器学习作为人工智能的关键核心技术,受到了前所未有的重视和快速发展。基于机器学习方式识别集装箱箱号,需要足够的集装箱箱号数据集,目前在港口领域,还没有成熟的集装箱箱号数据集。因此研究集装箱箱号图像样本扩充方法对于基于机器学习的集装箱箱号识别研究有着重要意义。本文主要研究内容如下:(1)集装箱箱号字符样本生成对抗网络(C-SAGAN)针对港口现场收集到的集装箱箱号图像数量较少,不足以用做识别训练的样本的问题,在分析集装箱箱号特征的基础上,设计一种用于集装箱箱号字符图像生成的生成对抗网络C-SAGAN。先将收集到的集装箱箱号图像中的字符划分成36种数据集,然后将这些数据集作为C-SAGAN模型的训练数据集。C-SAGAN基于SAGAN网络,将自注意力机制加入到生成器和判别器中,再加入字符的类别标签作为条件信息,以实现集装箱箱号字符图像样本的生成。实验结果表明,C-SAGAN生成的字符图像样本比CGAN生成的质量要更好。(2)基于Text Boxes++的集装箱箱号定位方法针对集装箱箱号的排序不唯一、箱体表面凹凸不平等的问题,设计一种基于Text Boxes++的集装箱箱号定位方法。采集的集装箱箱号图像中只有一个集装箱箱号出现,这便于箱号的定位。本文基于Text Boxes++算法上进行改进,把原生模型中的VGG网络更换成能够解决深度网络退化问题的残差网络(Res Net),再根据集装箱箱号的排列情况指定默认框的长宽比,进而快速定位箱号。实验结果表明,该算法相较于现有的集装箱箱号定位算法,有着较高的准确率。(3)基于CRNN的集装箱箱号识别方法针对集装箱箱号识别的问题,本文对CRNN卷积循环神经网络进行改进。CRNN是由CNN、Bi LSTM和CTC结合而成的网络,本文在CRNN中引入Bi GRU网络替换Bi LSTM网络,GRU网络保持了LSTM网络的效果同时又使结构更加简单,降低训练的难度。实验结果表明,本文集装箱箱号识别方法比现有的基于人工智能的识别方法在准确率上有所提升。更多还原

【Abstract】 As port transportation becomes more and more developed,in order to meet the needs of the development of port container business,the degree of automation of container operations is also getting higher and higher.The correct and efficient identification of container numbers is the basis of automated operations.As artificial intelligence is widely used in the port field,machine learning,as a key core technology of artificial intelligence,has received unprecedented attention and rapid development.Recognizing container box number based on machine learning methods requires sufficient container box number datasets.Currently,there is no mature container box number dataset in the port field.Therefore,it is great significance to research container image number expansion methods for container box number recognition based on machine learning.The main research contents of this thesis are as follows:(1)A generating adversarial network(C-SAGAN)for the generation ofcontainer box number characters.Aiming at the problem that the number of container box number images collected at the port site is small and not enough to be used as a sample for recognition training,on the basis of analyzing container box number characteristics,a generative adversarial network C-SAGAN for the generation of container box number character images is designed.First,the characters in the container box number image collected are divided into 36 datasets,and then these datasets are used as the training dataset of the C-SAGAN model.Based on the SAGAN network,C-SAGAN adds a self-attention mechanism to the generator and discriminator,and then adds the category label of the character as condition information to realize the generation of the container box character image sample.The experimental results show that the character image samples generated by C-SAGAN are better quality than CGAN.(2)Container box number positioning method based on Text Boxes ++Aiming at the problems of inconsistent ordering of container numbers and unequal unevenness on the surface of the boxes,a positioning method for container numbers based on Text Boxes ++ is designed.Only one container number appears in the collected container number image,which facilitates the positioning of the container number.This thesis improves on the Text Boxes ++ algorithm,replaces the VGG network in the native model with a residual network(Res Net)that can solve the problem of deep network degradation,and then specifies the aspect ratio of the default box according to the arrangement of container numbers,and then quickly locate numbers.The experimental results show that the algorithm has a higher accuracy than the existing container number positioning algorithm.(3)Container number identification method based on improved CRNNAiming at the problem of container number identification,this thesis improves the convolutional recurrent neural network(CRNN).CRNN is a network composed of CNN,Bi LSTM and CTC.In this thesis,the Bi GRU network is introduced in CRNN to replace the Bi LSTM network,the GRU network maintains the effect of the LSTM network while making the structure simpler and reducing the difficulty of training.The experimental results show that the container number identification method in this thesis is improved in accuracy compared with the existing artificial intelligence-based identification method.更多还原