Implementacija nizko-kompleksnega algoritma za brezizgubno kompresijo slik na FPGA

Regoršek, Žan (author), Trost, Andrej (mentor), Gorkič, Aleš (comentor)

V magistrskem delu obravnavamo algoritem za brezizgubno kompresijo slik in ga implementiramo na vezju FPGA. Slike so posnete v Bayerjevem vzorcu zato algoritem temelji na principu kompresija-najprej s čimer se izognemo potrojitvi količine podatkov že pred kompresijo. Z odstranitvijo barvne in prostorske odvečnosti zmanjšamo ponavljajočo informacijo. Sliko dodatno skrčimo z entropijskim kodiranjem, kjer uporabimo Golomb-Rice kodiranje z dinamično spremenljivim parametrom k (AGOR). Ker slikovni senzor pošilja po 16 podatkov hkrati, algoritem ustrezno paraleliziramo s čimer 16-kratno pospešimo kompresijo. Razvijemo lasten algoritem za vzporedno sestavljanje izhodnih kod, ki predstavljajo kodirano sliko. Raziščemo vpliv parametra k na učinkovitost ter izpeljemo pravilo za idealno vrednost parametra k, ki omogoči krčenje vse do teoretične limite. Programski model kompresorja in dekompresorja programiramo v jezikih C++ in Matlab. Na KODAK-ovi slikovni zbirki potrdimo pravilnost in ocenimo učinkovitost algoritma. V jeziku VHDL razvijemo vezje za strojno kompresijo slik, ki ga simuliramo s testnimi vektorji ustvarjenimi s programsko implementacijo algoritma. V evaluaciji rezultatov na KODAK zbirki slik ugotovimo, da algoritem slike v povprečju skrči na 75% izvirne velikosti, medtem ko slike iz naročnikove zbirke skrči tudi do le 44% izvirne velikosti. Na koncu predlagamo ideje za izboljšanje kompresijskega razmerja.

• Language: Slovenian
• Keywords: Golomb-Rice, kompresija, dekompresija, kompresija-najprej, DPCM, barvni prostor, FPGA, Bayer CFA, kapsulna endoskopija, AGOR
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2023
• PID: 20.500.12556/RUL-150172
• COBISS.SI-ID: 165255683
• Publication date in RUL: 14.09.2023

Secondary language

• Language: English
• Title: Implementation of a low complexity algorithm for lossless image compression on FPGA

Abstract

In this thesis Bayer CFA image lossless compression algorithm is evaluated and implemented on FPGA. Compression-first method is utilised to avoid demosaicing process which triples the original data. Image data is compressed through removal of color and spatial redundancy. It is further compressed by utilising Golomb-Rice entropy encoding with dynamically adjustable parameter k (AGOR). As the image sensor supplies data for 16 pixels simultaneously, the algorithm is paralelized, effectively making it 16 times faster. A way to assemble and concatenate bitstreams of compressed data is developed. The effect of parameter k on compression efficiency is studied and a rule for the ideal value of k is derived. The software implementation of the algorithm is programmed using C++ and Matlab. KODAK image dataset is used to verify the proposed algorithm and evaluate it's performance. VHDL is used to decribe a hardware implementation of the proposed algorithm. The design is tested and verified through simulation using test vectors generated by software implementation in C++. In performance evaluation it is found that the proposed algorithm in average compresses KODAK dataset images down to 75% of their original size, while high resolution images from another dataset are compressed down to 44% of their original size. In the end, some ideas are suggested as a way to further improve the compression ratio.

• Keywords: Golomb-Rice, compression, compression-first, DPCM, parallel implementation, FPGA, Bayer CFA, capsule endoscopy, AGOR, low complexity