Convolution neural network image classification method based on sparse coding pre-training

The invention discloses a convolution neural network image classification method based on sparse coding pre-training, At first, that train sample is transformed by non-down sample contourlet transform, The first two decomposed images are selected to expand the training samples, and then randomly selected images are used to learn their local features by SC algorithm, and the features are sorted according to the gray average gradient from large to small. Finally, the eigenvalues with larger gray average gradient are selected to initialize the CNN convolution kernel. The SC algorithm is used to initialize the convolution kernel of CNN by learning the statistical features of the original image, and the classification effect is better than that of the traditional bottom visual features, which effectively avoids the network training falling into local optimization. Synthesizing the advantages of high and low frequency sub-bands for different scenes, the image classification accuracy is effectively improved in the c