The cross-field DBN for image recognition

In this paper, a new cross-field deep belief network (DBN) model is designed to recognize images. The bottom layer of model is shared to capture low-level features while the high-layer neurons are detached to capture high-level features of the specific fields. The cross-field unlabeled sample sets are used to train the model due to the shared neurons, which can improve the recognition performance in current internet environment with large scale cross-field unlabeled images. The cross-field DBN is realized by Numpy and Theano, and tested by Mixed National Institute of Standards and Technology (MNIST) data set of handwriting characters and Columbia Object Image Library (COIL) data set of object images. The simulation shows that the cross-field DBN has better performance than traditional DBN due to the help of cross-field unlabeled training data. In addition, the training of sample set after binarization can greatly improve recognition performance.