Team develops image deep learning technology to present VR and AR screens more vividly and realistically - Tech Xplore

Professor Jin Kyong-hwan's probe squad of the Department of Electrical Engineering and Computer Science astatine Daegu Gyeongbuk Institute of Science and Technology (DGIST) has developed representation processing heavy learning exertion that reduces representation velocity and increases solution by 3dB compared to existing technologies.

Developed done associated probe with Choi Kwang-pyo of Samsung Research, this exertion reduces aliasing improvement connected the surface compared to existing awesome processing-based representation interpolation exertion (Bicubic interpolation), frankincense producing much earthy video output. In particular, it tin reconstruct the high-frequency portion of images clearly. It is expected to show a earthy surface erstwhile utilizing VR oregon AR.

Signal processing-based representation interpolation exertion (Bicubic interpolation) preserves desired images successful assorted environments by designating a circumstantial determination of an image. It has the vantage of redeeming representation and speed, but deteriorates the prime and deforms the image.

To code this issue, heavy learning-based, ultra-high-resolution video representation conversion technologies were developed, but astir of them are convolutional artificial quality network-based technologies, which person the disadvantage of inaccurate estimation of values betwixt pixels, which tin pb to representation deformation. Implicit look neural network exertion to flooded these disadvantages is drafting attention, but the disadvantage of implicit look neural web exertion is that it cannot seizure high-frequency components, and it increases representation and speed.

Professor Jin Kyong-hwan's probe squad developed a exertion that resolves the representation into respective frequencies truthful that the characteristics of high-frequency components tin beryllium expressed successful the image, and reassigns coordinates to resolved frequencies utilizing implicit look neural web exertion truthful that the representation tin beryllium shown much clearly.

It tin beryllium described arsenic a caller exertion that combines Fourier analysis, which is an representation heavy learning technology, and implicit look neural web technology. The recently implemented exertion tin amended implicit look neural networks that could not reconstruct high-frequency components by resolving indispensable frequence components successful restoring images done an artificial quality network.

Professor Jin Kyong-hwan said, "The exertion developed this clip is excellent, arsenic it shows higher restoration show and consumes little representation than exertion utilized successful the existing representation warping field. We anticipation that the exertion is utilized for representation prime restoration and representation editing successful the aboriginal and anticipation that it volition lend to some academia and industries."

Provided by DGIST (Daegu Gyeongbuk Institute of Science and Technology)