Digital recognition of street view house numbers based on DCGAN
Juping Zhong, Jing Gao, Rongjun Chen, Jun Li · 2019
Deep learning algorithms have surpassed human resolution in applications such as face recognition and object classification. However, it can only produce very blurred, lack of details of the image. Generative Adversarial Network is a game training of minimax antagonism between generator G and discriminator D, and ultimately achieves Nash equilibrium. We use deep convolutional GAN that recognizes sequence numbers and without split characters. First we use convolution network to extract character features. Second we construct a convolution neural network to recognize digits of natural scene house number. DCGAN is used to improve the resolution of the number of fuzzy houses, so as to extract more abundant data features in data set training. It can better recognize the numbers in the natural street.