Research And Application Of Text Recognition Algorithm In Complex Scenes

The Instrument,an important tool for detecting and recording the working status and historical data of the equipment,may cause “equipment island” in industry and life,making it prone to fail to monitor the devices online as a result of some problems,such as: the old-dated devices or an inconsistency in data interface.In this regard,this dissertation is inspired by a municipal program: the 2020 Zhengzhou Collaborative Innovation Major Special Project-Intelligent Industrial Internet of Things Platform Key Technology Research Project.It studies the instrument numbers recognition based on the text detection and recognition arithmetic to improve its accuracy in text recognition under the complex contexts,such as: background complexity,exposure transition,and dark environment.The dissertation elaborates as follows:(1)Regarding to the problem of text occlusion in complex scenes,which affects the accuracy of text region detection,this dissertation put forward a text region detection algorithm MF-RCNN(Multi-scale Fussion for Region CNN)by referring to the multi-scale fusion.The algorithm uses the residual network to extract image features,and adds multi-scale fusion modules FPN-I(FPN-Inception)into the convolution layer,making it possible to fuse the semantic information of high and low features to accurately identifying the location of small targets.The experimental results suggest that the algorithm levels up its classification precision to 89.4%,and its mean average precision to 58.4%.(2)As for a low text recognition accuracy resulted from some complex scenes,such as: noise and too dark light,this dissertation combines the MF-RCNN algorithm,and uses six data enhancement methods to expand the data to establish the related models to improve recognition accuracy.Based on it,the dissertation put forward another attention-based mechanism’s text recognition algorithm CS-CRNN(Convolutional Recurrent Neural Network with Channel).The algorithm incorporates an attention-based mechanism into the convolutional layer to improve the accuracy of identifying texts with large changes by linking context information.The experimental results suggest that the algorithm helps boost the accuracy in terms of the scroll wheel and nixie tube datasets respectively reaches up to 87.1% and 83.8%.(3)The dissertation develop a set of system for intelligent recognition text on the instrument,based on a instrument image arithmetic for text detection and recognition.It functions as a auto-recognition system for image acquisition,image processing,image text recognition,also a provider or corrector of the relative results with high speed and accuracy.To sum up,this paper conceives a algorithms of text detection and recognition for complex scenarios,and it helps develop a set of system for for intelligently recognizing industrial instrumentation data so that it can serve as a powerful auxiliary monitoring tool in the field of industrial instrument.