New approach for improving the image resolution using artificial intelligence

A new deep convolutional model for super-resolution of faces, developed by researchers from the University of Ljubljana Faculty of Electrical Engineering and the University of Notre Dame in the US, can improve the resolution of facial images as much as 8 times.

Authors: Klemen Grm, Vitomir Štruc, Walter Scheirer

Image resolution is of vital importance for human perception of image content and paramount for the success of various high-level computer vision tasks, such as object detection, target tracking, visual recognition or semantic scene understanding.

Researchers from the Laboratory for Machine Intelligence (LMI) of the University of Ljubljana Faculty of Electrical Engineering (Klemen Grm, Vitomir Štruc) and from the Computer Vision Research Lab (CVRL) of the University of Notre Dame in the US (Walter Scheirer) developed a new convolutional model for face super-resolution which boasts the capacity to improve the resolution of facial images as much as 8 times. As shown in Figure 1, the developed model is capable of generating convincing, high-resolution reconstructions even from extremely small images, in sizes of just 24 × 24 image pixels. At the core of the developed procedure is a new approach to training super-resolution models, which learns the model parameters using a loss function composed of an image reconstruction term and an additional term that relates to face recognition performance.

The model was created as part of the doctoral research of junior researcher Klemen Grm, and in the middle of September 2019 received the EAB Max Snijder award. The award is one of three awards presented by the European Association for Biometrics (EAB) each year to the best PhD theses associated with the field of biometrics in Europe. The paper with a description of the model has been published in a category A journal (very high quality achievements) – IEEE Transactions on Image Processing (TIP).

Source: Grm K., Scheirer W. J., Štruc V. Face hallucination using cascaded super-resolution and identity priors. IEEE Trans. Image Process., vol. 29, pp. 2150 – 2165, (2020), doi: 10.1109/TIP.2019.2945835.

Nov pristop k izboljšanju ločljivosti slikovnih podatkov z uporabo umetne inteligence

Examples of improving image resolution with the developed super-resolution model. The two images on the left show examples of input images for the developed model, the images in the middle show the model output, and the images on the right show the ground truth (or reference) high-resolution images intended for evaluating the results of the model.
Illustration: Klemen Grm, LMI, UL FE