The invention discloses a convolutional neural network algorithm optimization method based on a spare autocoder and a gray correlation analysis method, and belongs to the technical field of computer simulation. The method comprises the following steps that: firstly, through the spare autocoder, carrying out unsupervised pre-training on an input image to obtain one group of convolution kernel initial value sets which reflect original input image features; then, utilizing the group of convolution kernel initial values to carry out a convolution and pooling operation on the original input image;and importing the gray correlation analysis method to calculate an association degree between the feature image, which is pooled for the last time, of a sub-sampling layer and a corresponding output result, setting a threshold value, removing feature image data with small association so as to enable the system to automatically select a hidden feature image which has a high influence on an identification result. By use of the method, a false identification rate can be lowered, and verification effectiveness is verified.