Adaptive Local Receptive Field Convolutional Neural Networks for Handwritten Chinese Character Recognition

The success of convolutional neural networks (CNNs) in the field of image recognition suggests that local connectivity is one of the key issues to exploit the prior information of structured data. But the problem of selecting optimal local receptive field still remains. We argue that the best way to select optimal local receptive field is to let CNNs learn how to choose it. To this end, we first use different sizes of local receptive fields to produce several sets of feature maps, then an element-wise max pooling layer is introduced to select the optimal neurons from these sets of feature maps. A novel training process ensures that each neuron of the model has the opportunity to be fully trained. The results of the experiments on handwritten Chinese character recognition show that the proposed method significantly improves the performance of traditional CNNs.