Abstract: In order to solve the problem that the deep convolutional neural networks are difficult to be trained due to vanishing gradients, a straight convolutional neural networks algorithm based on improving the methodology of batch normalization is proposed. Firstly, the activations of convolutional layers are normalized. Secondly, the normalized activations are restored by reconstructing parameters. Finally, the proposed algorithm is used to train reconstruction parameters. On three image datasets CIFAR-10, CIFAR-100 and MNIST, the classification accuracies of SCNN can archive 94.53%, 73.40% and 99.74% respectively, which significantly outperforms other deep neural networks algorithms. The proposed algorithm can effectively overcome the problem of vanishing gradients in traditional convolutional neural networks.