Street View Character Recognition

Аннотация

This paper explores new methods and techniques for digital recognition based on convolutional neural networks using the Street View House Number dataset (SVHN) collected by Google. The main works are (1) Explore the current research status of digital recognition techniques and convolutional neural networks. (2) Optimize the door number recognition method in terms of training parameters, data augmentation, and model stacking according to the characteristics of SVHN dataset. (3) Improved the best performing model in terms of small target detection, the detection accuracy of the dataset reaches 94.9% and the detection speed reaches 7.9ms per image.