House Number Recognition Method Based On CNN Cross-layer Feature Fusion

Characters are one of the most important ways to express information,which has been widely used in house number recognition,license plate number recognition,postal area code recognition and other fields.However,due to the interferences of abrasion,tilt,illumination and the diversity of characters,there are still many practical problems which affect the accuracy and efficiency of character recognition algorithms in natural scenes.As a common model of deep learning,convolutional neural networks have been successfully applied in the field of image analysis and understanding.This dissertation discusses new methods and technologies for character recognition based on convolutional neural networks on Street View House Number(SVHN)dataset.Our main works are as follows:(1)The current researches on house number recognition and convolutional neural networks are firstly summarized.Then the composition modules and training process of convolutional neural networks are explained.Finally,the network structures and characteristics of LeNet-5 model,AlexNet model,GoogLeNet model and ResNet model are analyzed.(2)A house number recognition method is proposed according to SVHN dataset using optimized convolutional neural networks.Firstly,the influences of model components on the over all performance is studied,including activation functions,pooling methods,filters,convolutional layers and weights initialization methods.Secondly,a convolutional neural network which obtains high accuracy on SVHN dataset is constructed.Thirdly,featture extraction and classification are executed using the proposed model to obtain recognition results.Experimental results show that the recognition rate of the constructed model on SVHN testing set is up to 93.8%,which is superior to other methods such as HOG,KM-SVM,CNN-HMM,CNN-SVM,D-DBN,DBN,SDAE and LeNet-5.(3)A method of house number recognition based on convolutional neural networks with weighted cross-layer feature fusion is proposed,considering the fact that traditional convolutional neural networks only sent the features of last layer into the Soft-Max classifier,which ignores the detailed information extracted by the frontal layers.The Principal Components Analysis(PCA)method is used to reduce the dimensions of features to be fused at first.Then corresponding weights are computed according to their contributions to recognition results.Finally,the fused features are input to the Soft-Max classifier to get a more satisfying recognition result.Experimental results on SVHN dataset indicate that the proposed method could be fully trained within 2.2 hours,and the recognition rates is increased to 95.6%.