Abstract: The classical CNN usually discards negative feature information.However,it impacts the performance of image classification.To solve the problem,the Supplement CNN model is proposed.Firstly,the feature maps from convolution layer are reversed and the advantages of the Leaky ReLU function are employed to preserve the negative feature information.Secondly,this can increase the propagation of the effective classification information.The additional feature information will affect weight updating by back propagation algorithm,which benefit image classification.Finally,we illustrate how the number of layers affects the Supplement CNN model by experiments.Compared with the classical CNN and its partial fraction expansion methods,experimental results demonstrate the effectiveness and efficiency of our model.