De-warping Algorithm with Limited Line Buffers

Abstract

Denne oppgaven tar for seg det gjennomgripende problemet med bildeforvrengning – nærmere bestemt trommelforvrengning – som hovedsakelig oppstår i vidvinkelkameraer utstyrt med fiskeøye linser. Slike forvrengninger utgjør betydelige utfordringer i datasynsapplikasjoner, noe som krever robuste korreksjonsalgoritmer for nøyaktig bildeanalyse og beslutningstaking. Dette arbeidet presenterer en omfattende studie og utvikling av en forbedret sanntids algoritme for forvrengningskorreksjon. Algoritmen, forankret i Brown Conrady-modellen, utnytter kjente kameraforvrengningskoeffisienter for å rekalibrere og rette forvrengte bilder. En iterativ tilnærming ble tatt i bruk for å optimere korreksjonskoeffisientene på en slik måte at det gir etter for minimumskravet til linjebuffer for maskinvarerealisering, med Structural Similarity Index (SSIM) og Peak Signal to-Noise Ratio (PSNR) som verdiene. Den forbedrede algoritmen overgår forgjengeren betydelig når det gjelder PSNR-verdi ved å vise en forbedring på opptil 34,27 %, noe som indikerer en overlegen nøyaktighet i pikselverdien. Maskinvarerealisering av korreksjonsalgoritmen ble oppnådd gjennom en SystemVerilog-basert design, rettet mot FPGA-implementering. Studien identifiserer også områder for fremtidig forbedring, som for eksempel innføring av interpoleringsmetoder for ytterligere å øke SSIM-verdier og optimaliseringsstrategier for å redusere ressursforbruket. På grunn av begrenset tid, var full implementering av maskinvaredesignet på et FPGA-kort utsatt, synteseresultatene gir et solid grunnlag for fremtidige forsøk. Forskningsresultatene gir ikke bare en forbedret metode for bildeforvrengning, men også mulighet for fremtidige forbedringer i maskinvaredesign for bildekorreksjonsapplikasjoner.

The fidelity of image data is of paramount importance in today’s technologically driven landscape, where applications such as autonomous vehicles, robotic vision, and medical imaging rely heavily on the accuracy of visual information. This dissertation addresses the pervasive issue of image distortion—specifically bar rel distortion—that arises predominantly in wide-angle cameras equipped with fish-eye lenses. Such distortions pose significant challenges in computer vision ap plications, necessitating robust correction algorithms for accurate image analysis and decision-making. This work presents a comprehensive study and development of an improved real time barrel distortion correction algorithm. The algorithm, rooted in the Brown Conrady model, leverages known camera distortion coefficients to re-calibrate and correct distorted images. An iterative approach was adopted to optimize the cor rection coefficients in such a way that yields in to minimum line buffer requirement for hardware realization, with Structural Similarity Index (SSIM) and Peak Signal to-Noise Ratio (PSNR) serving as the figures of merit. The enhanced algorithm significantly outperforms its predecessor in terms of PSNR value by showing an improvement up to 34.27%, denoting a superior fidelity in pixel value accuracy. On the contrary, taking SSIM value in to account, the proposed algorithm has the average value of 0.8323 where as the best performing predecessor has a value of 0.9210, leaving scope for further improvement. Hardware realization of the correction algorithm was achieved through a Sys temVerilog based design, targeting FPGA implementation. The synthesis results from Vivado Design Suite highlight a design that, while resource-intensive, shows promise in throughput, capable of processing images at a rate of 105 frames per sec ond if required. Comparisons with previous work illustrate the advanced nature of the proposed design, suggesting higher resource usage but enhanced performance potential. The study also identifies areas for future improvement, such as the introduction of interpolation methods to further increase SSIM values and optimization strategies to reduce resource consumption.Although due to a limited amount of time, the full implementation of the hardware design on an FPGA board was out of scope, the synthesis results provide a solid foundation for future endeavors. In conclusion, this dissertation contributes a algorithm with certain amount of nobility for barrel distortion correction with practical implications for real-time image processing systems. The research outcomes not only offer an improved method for image de-warping but also pave the way for future improvements in hardware design for image correction applications.