Research On OMR Recognition Method Based On TensorFlow Convolutional Neural Network Platform

Solving image recognition problems with deep learning neural network technology is a hot topic in research and application.In this paper,the deep learning neural network model is applied to optical mark reading and recognition(Optical Mark Reader,referred to as OMR),which solves the problems of low erasure and low recognition rate of light coating points in traditional identification methods,and reduces the labor labor.The paper uses the improved projection segmentation method to obtain the coating information.Based on the TensorFlow platform,a special optical marker recognition neural network model(Convolutional Neural Networks Mark Reader,referred to as CMR)is built.The research and practice of the actual OMR recognition problem are carried out,and the deep learning recognition of the answer sheet is realized.The survey found that the traditional OMR uses the reader device and combined with the online marking system to identify the answer sheet,but this method has some problems in the identification process.The main performance is that the weight of the filling is large,the material and quality of the pen are different,the background of the paper has a certain degree of ambiguity,the scanning offset,the random stain,the erasure is not clean,etc.These factors lead to incomplete partial coating information.Problems such as unsatisfactory background separation,resulting in a recognition rate significantly lower than the human eye recognition rate,and even serious cases such as misunderstanding and rejection.Aiming at the above problems,this paper proposes an optical marker reading and recognition method CMR based on convolutional neural network,and carries out research and practice with the help of TensorFlow platform.The main research contents of the thesis are as follows:(1)For the problem of obtaining the information of the training data,the paper improved the Hough algorithm,solved the tilt problem in the data picture,and combined the projection method to segment the image and segment the image,and convert the result of the preprocessing.The TFRecord training data format required for TensorFlow.(2)For the identification problem of OMR accidental coating,based on the traditional OMR recognition,the neural network model CMR is constructed by using the deep learning platform TensorFlow.(3)For the optimization problem of CMR convolutional neural network model,the paper combines the deep learning theory knowledge to improve the model weight.Based on the data set and data set,the L2 regularization is used to convolution layer and pooling layer.The parameters are optimized to prevent over-fitting.The paper uses the view tool TensorBoard to view the calculation process and parameter settings of the TensorFlow model to help supervise and adjust the model.The experimental results show that the accuracy of the deep-learning network model in this paper is as high as 99.993%.Compared with the deep confidence network model DBN with the same depth of model,it has better efficiency and accuracy,and thus is more practical.In short,the CMR model solves the problem of inaccurate classification of the unclean and fill-coated lightcoated points in the traditional OMR recognition,and saves the manpower input while improving the recognition rate of the marking.