Research On Automatic Recognition Of Electronic Instrument Based On Machine Vision

With the continuous development of science and technology and the further innovation of information technology,the traditional production and lifestyle are all shifting towards digitalization and intelligence.As an important tool for data monitoring and measurement,electronic instruments play an important role in industries such as industry,agriculture,forestry,animal husbandry,and fishing,and are of great value in ensuring normal,reasonable,and safe production and life.Electronic instruments have been more and more widely used due to their advantages of strong intuitiveness and high measurement accuracy.However,at present,only a small number of electronic instruments have the function of transmitting data through the network,which can realize the automatic acquisition of measured values.On many outdoor portable devices,the meter representation still needs to be read and recorded manually.For such a way of recognizing and recording instrument data through human eyes,it is often easy to make errors due to the heavy workload of human eyes,or the recording personnel may subjectively omit part of the accuracy and only record rough data,thus reducing the accuracy of data collection.Therefore,an automatic data entry system for electronic instruments can effectively solve these problems.With the popularization of mobile terminals such as mobile phones,the method of acquiring images has become more convenient and the image quality has become clearer.Therefore,by taking pictures of electronic instruments through mobile phones and other devices,and then automatically identifying the measured readings through image processing and other methods,automatic identification and entry of data can be achieved.This thesis combines deep learning technology and image processing technology to study the automatic recognition method of electronic instruments.The method mainly includes five steps: detection and extraction of the instrument screen,positioning of key points on the screen,horizontal correction of the screen,text detection and display number recognition.In the research of electronic instrument screen detection,the thesis improves the SSD algorithm and proposes a Multi-Conv-SSD algorithm based on multi-convolution feature encoding.In this thesis,the multi-convolution structure reduces the detection model parameters and improves the detection accuracy,and realizes the separation of the screen and the background.As for the key point positioning method,this thesis proposes a key point detection model based on full convolution,which realizes the positioning of the key points of the screen in the rotated and distorted images.Then this thesis performs perspective transformation on the screen area according to the identified key points to achieve the horizontal correction of the screen.For text detection,this thesis uses CTPNbased text detection algorithm,and constructs the corresponding data-set to train the model.In text recognition,it is experimentally verified that the use of time series models such as LSTM to extract weak semantic features in the display text will degrade the model.The thesis proposes a text recognition algorithm based on convolution and self-attention modules,which acts on sample areas that are more difficult to identify to improve the recognition performance.The thesis collects the images of the instrument in the natural scene and constructs a data-set.The experiment is designed according to the proposed methods and compared on the data-set.The experimental results prove that the automatic recognition algorithms of electronic instrument proposed in this thesis can better recognize the instrument display text.