Precise Localization and Recognition of Train Characters

To solve the problems of the non-compact boundary localization and the loss of small objects,a train number localization and recognition method especially for precise localization of small-scale targets is proposed.This method improves from CTPN in the text localization stage.We first use the VGG16 network to extract features and fuse multi-scale feature maps,which is beneficial to small text areas and then use the region proposal network to generate candidate areas,for classification and regression.A hard sample mining strategy,that is,to retain positive samples containing only part of the train number is designed in the classification process.A boundary-sensitive fine-grained text box regression strategy in the regression process is designed to ensure that the horizontal boundary is compact.Finally the candidate regions are connected to output the localization results.In the text recognition stage,a text recognition method based on attention mechanism is used.By testifying the train number detection data set in the Caffe environment,the results show that the train number localization method is 0.11 higher than the classic text localization method.The overall train number recognition F1 value is 0.81.