The explosion of deep learning has made image recognition easier. Similarly, has also made a lot of progress in image repair. AI's restoration of images allows us to see that after learning a large amount of data, AI seems to be able to generate memory and imagination for images, restoring missing, blurry, or noisy images to their "original appearance."
# Next, let’s take a look at the progress of various image repair technologies. (Recommended learning: PHP video tutorial)
No need to clean samples, super denoising
Recently, NVIDIA, Aalto University and Massachusetts The Polytechnic University together proposed a new technology for image restoration, which can effectively remove noise and artifacts from images and does not require clean image samples. The work was announced at the 2018 ICML conference.
The video shows different image noises (including Gaussian noise, Poisson noise, Bernoulli noise, impulse noise, etc.). The neural network achieves good results by learning pairs of noise pictures.
It is considered to be a very good person in this work at present, and some details are also handled quite well. They call this technology Noise2Noise. The team obtained 50,000 images from the ImageNet database and "noised" them. Then input these "dirty" pictures into the model for training, so that the model can learn to "reduce noise".
It is worth noting that all the pictures taken over by this model are pictures with various noises added, and there is no idea what the original picture looks like. The researchers said: "It is possible for neural networks to learn to recover images without clean images." So they used paired noisy images to complete the work.
The researchers hope to apply this technology to images containing a lot of noise, such as astrophotography, magnetic resonance imaging (MRI) and brain scans.
Use nearly 5,000 images from the IXI dataset to train Noise2Noise’s MRI image denoising capabilities. Without artificial noise, the result may be slightly blurrier than the original image, but still restores sharpness well.
AI brain supplement, repair missing images
There are many algorithms for repairing missing images. Let’s first look at the repair algorithm for missing entire images. This algorithm Jobs from the University of Southern California.
Although it doesn’t look perfect, as a PS newbie, I can only operate it to this level.
CNN network structure
The network is actually composed of two neural networks, one is a content generation network and the other is a texture generation network. A content generation network generates images and infers the content of missing parts. The texture generation network is used to enhance the texture produced by the content network. Specifically, the generated completed image and the original non-missing image are input into the texture generation network, and the loss is calculated on a certain layer of feature_map, which is recorded as Loss NN.
For more PHP related technical articles, please visit the PHP Graphic Tutorial column to learn!
The above is the detailed content of what is ai repair. For more information, please follow other related articles on the PHP Chinese website!