This technique uses a convolutional neural network to accomplish something similar, but in the end, it doesn't give us an image, but a set of convolution kernels.

这项技术用了卷积神经网络以完成类似的东西，但最终它并没有给出一副画面，而是一系列的卷积核心程序。

Since we have a neural network over here, it goes without saying that the training takes place on a large number of before-after image pairs, so that the network is able to produce these convolution kernels.

既然我们这里有了神经网络，这个训练理所当然地发生于大量的前后相邻的像对中，因此这个网络能够制成这些卷积核心程序。