Deep convolutional neural networks (CNN) have achieved the unwavering\nconfidence in its performance on image processing tasks. The CNN architecture\nconstitutes a variety of different types of layers including the convolution\nlayer and the max-pooling layer. CNN practitioners widely understand the fact\nthat the stability of learning depends on how to initialize the model\nparameters in each layer. Nowadays, no one doubts that the de facto standard\nscheme for initialization is the so-called Kaiming initialization that has been\ndeveloped by He et al. The Kaiming scheme was derived from a much simpler model\nthan the currently used CNN structure having evolved since the emergence of the\nKaiming scheme. The Kaiming model consists only of the convolution and fully\nconnected layers, ignoring the max-pooling layer and the global average pooling\nlayer. In this study, we derived the initialization scheme again not from the\nsimplified Kaiming model, but precisely from the modern CNN architectures, and\nempirically investigated how the new initialization method performs compared to\nthe de facto standard ones that are widely used today.\n