НИУ ВШЭ: LAPUSKA (LaPlacian UpScale Knowledge Alignment)

The main articles are:

• Neural networks (neural networks)
• Video analytics (terms, applications, technologies)
• Computer vision (machine vision)

2023: Presentation of the updated architecture of the LAPUSKA neural network

Scientists at MISIS University and HSE on November 29, 2023 presented an updated architecture of the LAPUSKA neural network (LaPlacian UpScale Knowledge Alignment), which can significantly improve the quality of images. The proposed approach allows you to process photos 2 times faster than the chosen counterparts. In the future, the development of researchers will help to recognize faces and more accurately process images.

Analysis and interpretation of images or videos with, machine vision is already applied in,,, and to medicine agriculture transport entertainment industry many other areas. A promising direction computer vision is the super resolution of images, which not only increases the size of the image, but also improves its quality. This allows you to see more information and details that were inaccessible to human vision at low photo resolution.

соавтор исследования Ilya Makarov, Director of the Center for Artificial Intelligence NUST MISIS, Head of the AI in Industry Group of the AIRI Institute

Ultra-high resolution imaging technology helps overcome the limitations of photo and video devices and can be useful in a variety of practical applications. For example, in the field of security, super-resolution of images helps to increase the quality of images from CCTV cameras for recognizing faces or car numbers, in the field of digital image processing - helps to restore old or damaged photos, as well as improve the quality of photos taken in difficult lighting conditions or over a long distance.

Existing models for ultra-high resolution imaging have significant drawbacks, such as the most neural SRGAN and LapSRN models require large computational costs and a significant amount of computer memory, which affects the availability of their use and the time required to obtain the result. LapSRN usually allows you to get smoother images, which leads to the loss of some small details, while in images processed with SRGAN, there is a lot of digital noise.

The updated LAPUSKA neural network architecture for high image resolution combines the best properties of existing SRGAN and LapSRN models and addresses their disadvantages. The proposed model has a quality similar to LapSRN, but it is more than 2 times faster in processing time.

The architecture of our proposed neural network consists of several convolutional layers with different structures. At the heart of the network structure is the SRGAN structure, which performed well during training, and uses a post-upsampling strategy in which features are extracted directly from the LR input by a set of residual blocks, and the image is scaled at the end of propagation. An important point in the learning process of implemented models is the training data. In this work, it was decided to use the DIV2K datacet, since it contains 800 RGB HR training color images with corresponding reduced LR images with different coefficients, "said study co-author Ilya Makarov, director of the NUST MISIS artificial intelligence center, head of the AI in Industry group at the AIRI Institute.