| Character recognition is one of the basic contents in computer vision and pattern recognition, which is helpful to the consequent information processing. Regarding the Street View House Number images as the object of study, we combine convolutional neural network with support vector machine (SVM) to explore new recognition methods on the Street View House Number.The main innovations of this dissertation are as follows:(1) An image preprocessing method is proposed to deal with the character image in real world. The method employs different RGB weights to get the grayscale images. Then, the grayscale images are sharpened with Laplacian processing to enhance the details in the images. Experimental results show that compared with the original image, the network structure can be stabilized more quickly when the preprocessed images are trained by the convolution neural network.(2) Focusing on the SVHN dataset acquired in real environment, a recognition method based on an improved LeNet-5 model is proposed for Street View House Numbers, which combines convolutional neural network with support vector machine. In this method, after original color images are transformed to grayscale images, some images enhancement are employed to highlight the number features. Then, the third convolutional layer is omitted and the Softmax classifier of the final output layer is replaced with SVM classifier in the LeNet-5 model, which not only simplifies the network construction, but also helps to obtain a higher recognition rate. Experimental results show that the recognition rate of the proposed method may reach 90.35% after the sample images were trained with only 7 hours, which efficiently improved the overall ability of the recognition.(3) On the basis of the above work, a recognition software for street doorplate number is designed and implemented. The software provides some functions such as network training, network testing and doorplate number recognition, whose GUI is simple and friendly. |
PDF Full Text Request