Abstract: Convolution neural network models require the consistency of spatial scales between training images and testing images. In order to alleviate this restriction, a scale invariant convolution neural network model with multi-scale feature extractor was proposed, which can adapt to the in-plane scale change of input images. Meanwhile, multi-layer Maxout networks are nested into the model in order to improve the ability of feature extraction, so as to improve the accuracy of image recognition and classification. Experiments show that the new model improves the scale invariance and classification accuracy of traditional convolution neural networks.

Key words: convolutional neural networks, scale-invariant, Maxout, deep learning