Shulou(Shulou.com)06/02 Report--

What this article shares with you is about the principle of AI face changing technology. The editor thinks it is very practical, so I share it with you to learn. I hope you can get something after reading this article.

Recently, AI face-changing technology has caused great controversy. As we all know, face exchange is a hot application in the field of computer vision, while the technological progress has also buried a lot of hidden dangers. Therefore, in the recent "regulations on the Management of Network Audio and Video Information Services", the state issued a management and control policy for AI face-changing and fake video. Thus it can be seen that the rapid development of AI face-changing technology has even reached the stage of confusing the real with the fake. Since AI technology is so amazing, let's take you to an in-depth analysis of the specific principles of AI technology.

The principle of AI face changing technology:

Face exchange can generally be used for video synthesis, privacy services, portrait replacement or other innovative applications. Before the earliest, face exchange was achieved by analyzing the similar information of the two faces, that is, through feature point matching to extract feature information from one face, such as eyebrows and eyes, and then match it to another face. This implementation does not require training, unneeded data sets, but the implementation is relatively poor, unable to modify facial expressions.

In the recently developed deep learning technology, we can extract the deep information of the input image through the deep neural network, so as to read out the hidden deep features to achieve some novel tasks, such as style transfer (style transfer) is to read the trained model to extract the deep information in the image to achieve style exchange. Neural network is also used for face exchange (face-swap), in which VGG network is used to extract features and realize face exchange. Here we use a special self-encoder structure to achieve face exchange, and achieve good results.

Self-encoder:

The basic background of AI face changing Technology-- self-Encoder

Self-encoder is similar to neural network, which can be said to be a kind of neural network. After training, it can try to copy the input to the output. Like the neural network, the self-encoder has a hidden layer h, which can parse the input into a coding sequence and reproduce the input. From inside the encoder, there is a function hcorrecf (x) that can be encoded, and a function riterg (h) that decodes it, as shown in the following figure.

Network architecture:

So how to realize our face-changing technology through self-encoder? We have known before that the self-encoder can learn the information of the input image to encode the input image information and store the encoded information in the hidden layer, while the decoder uses the learned hidden layer information to reproduce the previously input image, but what happens if we directly input the images of two different individual image sets into the self-encoder? If we simply throw a collection of two different faces into a self-coding network and choose a loss function to train, but we get nothing in this way, so we need to redesign our network.

How to design it? Since we want to exchange two faces, we can design two different decoding networks, that is, we can use one coding network to learn the common features of two different faces, and two decoders to generate them respectively. That is, we design an input or an encoder (input two different faces), and then two outputs or two decoders, so that we can generate two different faces through the hidden layer.

Network structure:

We can see that the network structure has one input and two outputs, the input is composed of convolution layer and full connection layer, and the output is also composed of convolution layer, but it should be noted that the input is downsampling convolution and the output is upsampling convolution, that is, the resolution of the image decreases first and then increases slowly.

Generally speaking, this face-changing technology is a small project with simple structure but rich knowledge. Its structure is simple, easy to use and modify, and can produce good results, but because it has more parameters, its running speed is not fast (of course, we can speed up the training generation by changing the structure of the coding layer and the coding layer). And images with foreign bodies on the face may have an unrealistic effect.

These are the principles of AI face changing technology, and the editor believes that there are some knowledge points that we may see or use in our daily work. I hope you can learn more from this article. For more details, please follow the industry information channel.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.