Journal of Digital Information Management Yuan Zhang, BiMing Shi 15 1 2017 At present, the traditional convolutional neural network (CNN) can easily cause overfitting in the training process, thereby resulting in an invalid training model. Thus, this study proposed a novel CNN regularization method to avoid overfitting in the training process and to increase the image classification accuracy of CNN. The proposed method uses failure probability as the theoretical basis. First, the failure probability density (FPD) function of image pixel points was deduced, and the FPD prediction model of pixel points in image neighborhood was established. Second, this method evaluated the FPD of image eigenvalues, selected the pixel points with the minimum FPD in the image neighborhood as the retention characteristics, and extracted excellent features. Finally, this method conducted a labeling classification experiment on three image classification datasets (MNIST, CIFAR-10, and CIFAR-100) and compared the image classification accuracy of the three existing popular pooling methods (Max, Average, and Stochastic pooling). Results demonstrated that the proposed regularization method based on FPD can increase the classification accuracy of the image data. The classification accuracies gained by the training datasets are all higher than 90%, and the classification accuracies gained by the testing datasets are all higher than 85%. The proposed CNN regularization method possesses good generalization capability and engineering applicability.