Abstract | U ovom radu će biti obrađeni pojmovi kao što je dubina bita i percepcija dubine. Biti će riječi
o neuronskim mrežama, dubokim neuronskim mrežama, metodama korištenja neuronskih mreža,
interpolaciji i tehnikama povećanja dubine bitova.
Neke od novijih metoda korištenja neuronskih mreža su BitNet, BE-CALF, BE-ACGAN i
BrNet neuronske mreže. S obzirom na sliku niže dubine bitova, zadatak povećanja dubine bitova
je nadopunjavanje manje značajnih bitova koji nedostaju prema zadanim više značajnim
bitovima te vratiti prikladnu reprezentaciju slike visoke dubine bitova. U radu će se opisati neke
klasične metode povećanje dubine bitova, poput nadodavanja nula, ponavljanja bitova, MRC,
ACDC i IPAD algoritama te različite vrste interpolacije kao korak ispred povećanja dubine
bitova.
Kao praktični zadatak korišena je aplikacija Matlab u kojoj je odrađena interpolacija dubina
bita iz 6-bitne u 8-bitnu te iz 8-bitne u 10-bitnu, 12-bitnu i 16-bitnu sliku. Rezultati su
uspoređeni korištenjem objektivnih mjera kvaliteta poput MSE, PSNR i SSIM, MSSIM mjera
(koje koriste referentnu sliku) te BRISQUE, NIQE ili PIQE (koje ne koriste referentnu sliku) kao
i ekspertnom analizom subjektivnog doživljaja. U istraživanju je također korištena metoda
množenja 8 bitne slike s omjerom x4 (točnije 1023/255), x16 ( točnije 4095/255) ili x257
(točnije 65535/255). Uz to sve je ispitano vrijeme izvođenja za mogućnost korištenja algoritama
u realnom vremenu, i u ovisnosti o rezoluciji slike.
Na samom kraju je iskazan objektivni i subjektivni zaključak na temelju svih mjera kvalitete i
prikazane sve dobivene vrijednosti i srednja vrijednost u tablicama za svaku pojedinu sliku u
.ppm formatu. Bitno je napomenuti da je neke tehnike bilo teško procijeniti na običnim
monitorima. |
Abstract (english) | The concepts of bit depth and depth perception will be addressed in this Master's thesis.
Topics such as neural networks, deep neural networks, methods of using neural networks,
interpolation, and techniques for increasing bit depth will be discussed. Some of the newer
methods utilizing neural networks include BitNet, BE-CALF, BE-ACGAN, and BrNet neural
networks. Given an image with lower bit depth, the task of increasing the bit depth involves
filling in the missing less significant bits based on the given more significant bits to restore an
appropriate representation of a high bit-depth image. The paper will describe some classical
methods of increasing bit depth, such as zero padding, bit repetition, MRC, ACDC, and IPAD
algorithms, as well as various types of interpolation as a step towards increasing bit depth.
As a practical task, the Matlab application was used to perform bit depth interpolation from
6-bit to 8-bit, as well as from 8-bit to 10-bit, 12-bit, and 16-bit images. The results were
compared using objective quality measures such as MSE, PSNR, SSIM, MSSIM measures
(which use a reference image), as well as BRISQUE, NIQE, or PIQE (which do not use a
reference image), along with expert analysis of subjective perception. The research also
employed the method of multiplying an 8-bit image by a factor of x4 (more precisely, 1023/255),
x16 (more precisely, 4095/255), or x257 (more precisely, 65535/255). Additionally, the
execution time was examined to assess the possibility of real-time algorithm usage and its
dependence on image resolution.
At the end, an objective and subjective conclusion is provided based on all quality measures,
and all obtained values and their means are presented in tables for each individual image in .ppm
format. It is important to note that some techniques were challenging to evaluate on regular
monitors. |