为了有效解决铁路计算机联锁系统人工测试效率低下、准确性低等问题,提出了一种面向铁路联锁上位机界面文本定位与识别的深度学习模型。首先,通过对比基于CTPN(Connectionist Text Proposal Network)神经网路的文本定位模型与ResNet50、AlexNet、ZF以及VGG16四种卷积神经网络分别结合的策略,选择了最优的特征提取网络VGG16,增强了卷积特征图表示高层语义信息的细节特征,以利于定位小文本区域。其次,对比常见的目标检测模型在文本定位效果上的表现以及采用dropout方法,提升了文本定位模型CTPN网络的泛化能力和鲁棒性。然后,采用水平投影和垂直投影相结合的分割方法,可有效避免联锁上位机界面文本黏连等问题。最后,采用改进的AlexNet网络,输出文本识别结果。通过在TensorFlow环境验证联锁上位机界面文本定位和识别数据集,结果显示,CTPN网络在铁路联锁上位机界面文本定位精确率上达到了87.98%,召回率73.33%,调和平均数指标80.39%；改进的AlexNet网络文本识别准确率达到了89%。说明本文方法能够对铁路计算机联锁上位机界面文本实现准确定位和识别,并为联锁自动测试过程中自动办理进路及测试结果自动分析提供可靠的数据支持。

In order to effectively solve the problems of low efficiency and low accuracy of manual testing of railway computer interlocking systems, a deep learning model for text localization and recognition of the upper computer interface of railway interlocking is proposed. First, by comparing the strategy of combining the text localization model based on CTPN (Connectionist Text Proposal Network) neural network with the four convolutional neural networks of ResNet50, AlexNet, ZF and VGG16, the optimal feature extraction network VGG16 was selected, which enhanced Convolutional feature maps represent the detailed features of high-level semantic information to facilitate the location of small text areas. Secondly, comparing the performance of common target detection models on the text localization effect and adopting the dropout method, the generalization ability and robustness of the CTPN network of the text localization model are improved. Then, the use of a combination of horizontal projection and vertical projection can effectively avoid problems such as text sticking on the interface of the interlocking upper computer. Finally, the improved AlexNet network is used to output the text recognition results. By verifying the text positioning and recognition data set of the interlocking host computer interface in the TensorFlow environment, the results show that the accuracy of the text positioning of the CTPN network in the railway interlocking host computer interface has reached 87.98%, the recall rate is 73.33%, and the average index of harmony is 80.39%; The improved AlexNet network achieves 89% accuracy in text recognition. It shows that the method in this paper can accurately locate and recognize the interface text of the railway computer interlocking upper computer, and provide reliable data support for automatic routing and automatic analysis of test results during the automatic interlocking test process.

TP391.1

甘肃省科技计划项目(20CX9JA125,20JR5RA407)