Deep Learning Based Map Code Recognition Method

Abstract

For the efficient digitalization of maps, it is neces-sary to register map code, which indicates scale and area range. The manual register of the map code is time-consuming and laborious. Moreover, the recognition accuracy of the traditional text detection method based on template matching cannot be guaranteed. In this study, we propose a deep learning method to recognize image code and solve the problem of inaccurate detection and inaccurate positioning of text areas in map code recognition. A map code target recognition neural network referred to as YOLO-MCRNN is trained first to identify, crop, and binarize the code area and obtain detailed images. Architec-turally, an attention mechanism is added to a bidirectional long short-term memory structure and integrated into the map code target recognition network. Single numbers were cut according to variations in pixel value to establish a template library used in the matching method to recognize the code. Then, we conducted a comparative experiment to demonstrate the efficacy of the proposed deep learning method, and the results show that the detection network model improved on the performance of the existing state-of-the-art methods. Recognition experiments were carried out on the image dataset of code images taken by a digital map instrument. The single character and code recognition accuracy of the proposed YOLO-MCRNN model were 0.956 and 0.985, respectively, exhibiting obvious advantages over conventional matching methods and the unimproved text detection model CRNN(Convolutional Neural Network). The YOLO-MCRNN map code recognition network can accurately locate text areas, extract effective deep features of an input image, and identify map code. Compared with CRNN neural networks, which lack regional positioning, YOLO-MCRNN can improve recognition accuracy and reduce time consumption. Thus, the proposed model can be highly effective as an automatic map code registration method.