基于TensorFlow平台的卷积神经网络对OMR识别问题研究

Study on the OMR recognition based on convolutional neural network based on TensorFlow platform

文章针对光学标记阅读(OMR)识别技术中存在的背景虚化和复杂操作导致加大人力劳动的问题,提出一种基于TensorFlow平台的卷积神经(CNN)模型架构。以提高OMR识别技术为目的,涵盖深度学习开源工具TensorFlow与神经网络研究领域,以基于TensorFlow框架构建卷积神经网络为研究对象,运用TensorFlow学习系统构建神经网络模型,实现对OMR识别问题的研究。文章利用TensorFlow对OMR输入数据进行预处理,得到TFRecord文件,再在CNN神经网络架构基础上搭建训练模型实现对OMR答题卡数据的识别,得到标准数据。并运用视图工具TensorBoard来有效显示TensorFlow模型在运行过程中的计算流图以及模型参数随着训练的变化。

Aiming at the problem of increasing the labor force caused by background blurring and complex operations in optical mark reading(OMR)recognition technology, a convolutional neural network(CNN)model architecture based on TensorFlow platform is proposed. For the purpose of improving OMR recognition technology, it covers the field of deep learning open source tools TensorFlow and neural network research. The TensorFlow framework is used to construct convolutional neural network as the research object. The TensorFlow learning system is used to construct the neural network model to realize the research on OMR recognition. The paper uses TensorFlow to preprocess the OMR input data, obtains the TFRecord file, and builds the training model based on the CNN neural network architecture to realize the identification of the OMR answer sheet data and obtain the standard data. The view tool TensorBoard is used to effectively display the calculated flow graph of the TensorFlow model during operation and the changes of the model parameters with the training.