LAUSR

In order to apply deep learning to the field of image recognition, the basic principle, training process and model structure of deep belief networks (DBNs) in deep learning are analysed. For small samples, samples are down-sampled at the pretraining stage. In the parameter fine-tuning stage, random dropout is introduced, and the hidden layer nodes are cleared out randomly to keep the weights unchanged. The results show that the layered training mechanism of DBNs greatly reduces the difficulty of training and the training time. In the small sample, after introducing the down-sampling and random dropout, the deep belief network has a good improvement in recognition rate and time consumption, and the over fitting phenomenon is effectively alleviated.