基于全卷积网络的燃气表数字的分割与识别

Segmentation and Identification of Gas Meter Number Based on Full Convolution Network

随着计算机性能的逐渐提高,深度学习得到了广泛关注。通过低层的特征组合得到更为抽象的高级特征被广泛应用于计算机视觉、模式识别等领域。卷积神经网络(CNN)主要通过对二维图形的识别来实现图像级别的分类,全卷积网络(FCN)在CNN的基础上还具有全卷积化、上采样和跳跃结构的特点。笔者提出了一种基于全卷积网络的燃气表数字分割与识别方法,首先对数据集进行自主标注,然后在提出的FCN网络结构基础上进行燃气表数字分割,最后在预训练模型的基础上进行二次训练,实现了燃气表的数字识别。实验结果表明,该数字识别模型能够对清晰的数字进行准确识别;但是对于模糊数字和不完整数字的识别精准度还有待提高。

With the gradual improvement of computer performance,deep learning has received more and more attention.Because it explores the abstract high-level features by a combination of low-level features,deep learning has been widely used in various fields,such as computer vision and pattern recognition.Convolutional Neural Networks(CNN)mainly was used in recognition of images and it can achieve the classification in image level.FCN has the characteristics of fully convolutional,up-sample and skip architecture on the basis of CNN.This paper introduces the FCN that can predict the segmentation and digital recognition on gas mater.The method labels the data set independently,and then gives the network structure diagram.Simultaneously,we adopt second training based on the pre-training model.Finally,a simple digital recognition model is implemented.The results show that this model does well in recognizing clear figures,but it can't accurately recognize fuzzy or incomplete figures.