基于深度卷积神经网络的街景门牌号识别方法

A Method of Street View Number Recognition Based on Deep Convolution Neural Network

在复杂图像中提取文本信息是模式识别研究的热点,应用前景广阔。自然场景中的门牌号背景复杂,字符风格多样,识别难度较大。基于卷积神经网络设计一种识别方法,可以达到较好的识别效果。在方法设计中用灰度化手段来弱化自然场景中的背景信息,突出重要特征。基于AlexNet改进网络,加深网络的深度,在激活函数的后面使用批归一化BN,并在全连接层中应用较低比例的Dropout策略。使用谷歌街景门牌号数据集(SVHN),训练约13个小时,识别率达到94.58%。

Extraction of text information from complex images is a hot topic in pattern recognition and has a wide application prospect.In the natural scene,the house number is not only complicated in background design,but also in character style,which is difficult to recognize.Designs a character recognition method of Street View house number based on the convolution neural network,which can achieve better recognition effect.In the design of the method,gray preprocessing method is used to weaken the background information in the natural scene and highlight the important features.Improve network based on AlexNet,deepen network depth and optimize network structure.BN(Batch Normalization)is used after activation function,and a lower scale Dropout policy is applied in the full connection layer.Using Google Street View house number Data Set(SVHN),the training lasts approximately 13 hours and the recognition effect is 94.58 percent.