Multi-digit Recognition Of Natural Scene House Number Based On Deep Learning

It has always been research hot spots and difficulties in the field of computer vision that the desired character information is extracted from the complex graphic images.The natural scene number is severely distorted due to blurred images,uneven illumination,weak illumination,etc.,which makes it difficult to achieve an ideal personality recognition effect,and it is more difficult to recognize characters of any length.So,this thesis takes the image of the house number photographed in the real scene as the research object,and conducts in-depth research on the convolutional neural network(CNN)and the generated confrontation network(GAN).The main research work is as follows:First of all,a single digital identification network is constructed based on convolutional neural network for the problem that the current single digital house number recognition rate is not high and the algorithm is complex.Firstly,a single digital identification network consisting of input layer,two layers of convolutional layer,two layers of pooling layer,full connection layer and output layer is constructed.The convolutional neural network is used to automatically extract image features.In order to highlight important features,gray-scales were used to weaken background information in natural scenes when designing methods,and a certain proportion of Dropout strategies were utilized to prevent overfitting.Finally,by verifying on the SVHN dataset,the single digital house number recognition rate reached 95.72%.Compared with the algorithm results of other articles,the method is superior to most of the existing algorithms.Secondly,a deep convolutional neural network capable of simultaneously identifying multiple digits is constructed for the problem that the character segmentation is large and easy to cause errors.Based on the single digital house number identification network,a circular network is used to generate character sequences,and a 12-layer deep convolutional neural network is constructed by convolving the convolutional neural network.In the case of not dividing characters,the deep convolutional neural network can recognize multiple digits at the same time and perform verification on the SVHN data set,and achieves good results.Thirdly,aiming at the problem of recognition error caused by the blurring of the image dataset in the simulation result analysis of the house number identification,this paper uses the deep convolution generator to generate new high-resolution samples against the network,and extracts more abundant data in the dataset training process,and improve the effect of recognizing fuzzy house number.Finally,On the basis of this,I designed house number recognition Application software.The software interface is simple and user-friendly.It mainly has two functions of selecting photos and identifying.It can select the required house number picture from the album to identify.