The invention discloses a deep neural network acceleration method and apparatus. The method comprises the following steps of S1, defining a wavelet-like auto-encoder, and performing learning of the wavelet-like auto-encoder to obtain network parameters of the wavelet-like auto-encoder; and S2, performing image classification based on the wavelet-like auto-encoder. The purpose of accelerating a deep neural network can be achieved on the basis of not performing explicit modification on the network, so that multiple different network structures can be generalized.