A Real-Time FPGA-Based Architecture for a Reinhard-Like Tone Mapping Operator

Presentation Transcript

A Real-Time FPGA-Based Architecture for a Reinhard-Like Tone Mapping Operator Firas Hassan and Joan Carletta The University of Akron

Outline of Presentation Background and objectives Existing strategies for neighborhood tone mapping Real-time minor departure from the Reinhard administrator Experiments and results Future work

Outline of Presentation Background and objectives

Luminance and element run Luminance relate to pixel power Different gadgets are delicate to various scopes of luminance: Human visual system: 14 log units or 48 bits Imaging sensors: 9.5 log units or 32 bits Conventional displays 2.5 log units or 8 bits Mismatch in element ranges makes it: Hard to catch scenes as human see them Even harder to show these scenes!

Tone mapping administrators are utilized to connect the bungle between HDR pictures and show gadgets pack the dynamic scope of HDR pictures to displayable range repeat however much as could be expected of the visual impression of the scene

Global tone mapping administrators are free of neighborhood spatial setting perform same operation on every pixel don't function admirably when brightening fluctuates locally

Local tone mapping administrators shift adaptively with the nearby attributes of the picture create higher quality tone mapped pictures than worldwide TMOs can require complex calculation can experience the ill effects of radiance curios

Goals of research Develop calculations for nearby tone mapping of dark scale HDR pictures to such an extent that: they can be appeared with clear detail on standard showcases handling is "ongoing" (60 outlines/second for standard LCD screens) preparing can be effortlessly installed (utilizing field programmable door clusters) The framework is the consequence of a watchful exchange off of both picture preparing and equipment execution perspectives

Outline of Presentation Existing techniques for nearby tone mapping

Basic structure of nearby TMOs L: luminance I: enlightenment, identified with lighting conditions R: reflectance, identified with question in scene

Retinex strategy (Jobson et al.) utilizes Gaussian encompasses focused on a pixel to gauge nearby brightening single-scale: utilize one settled size encompass (get corona antiquities) multi-scale: utilize mean of three diversely measured encompasses standardizes every pixel by its nearby enlightenment distributed 2004 usage is not continuous: single-scale Retinex 256 × 256 grayscale picture 20 outlines/sec on an advanced flag processor

Reinhard technique utilizes the best brightening gauge around the pixel from a Gaussian pyramid of the picture wipes out radiance ancient rarities superior to anything Retinex distributed 2005 execution is not constant: utilizes four-scale Gaussian pyramid 1024 × 768 shading picture 14 outlines for each second on representation card bottleneck was memory data transmission

Reinhard technique Reinhard's choice of best window for light gauge too little an encompass gives poor gauge (differentiate misfortune comes about) too expansive an encompass may incorporate light source (coronas result) begin with littlest encompass consider next biggest encompass; if its normal is very little not the same as the littler surround's, utilize it rather result: utilize greatest encompass that doesn't contain a major change in enlightenment

Other neighborhood TMOs All can dispose of radiance relics yet too complex to be in any way continuous! Iterative techniques Low ebb and flow picture simplifier Gradient area HDR pressure Nonlinear channels Bilateral channels Trilateral channels Image appearance models Pattanaik iCAM

Outline of Presentation Real-time minor departure from the Reinhard administrator

Block chart of our strategy

Approximating a Gaussian surroud falling geometric arrangement rising geometric arrangement

Implementing the window with aggregators

Four-scale estimated Gaussian pyramid

Hardware for the 56x56 pixel window Horizontal Computation Block (HCB) Vertical Computation Block (VCB) Complete Hardware

Memory association

Log normal

Normalize the pixel Fast equipment for the complementary maintain a strategic distance from division – it's costly and moderate! acquire from the Newton-Raphson calculation, iteratively discovers foundations of a capacity base of is corresponding of b calculation says here; this implies look into introductory figure in view of 8 bits of mantissa of b ; one cycle then gives 17 bits of proportional

Outline of Presentation Experiments and Results

Simulation comes about

Simulation comes about

Simulation comes about

Hardware combination comes about Compatible with a casing rate of 60 edges/sec (for not exactly a full-sized LCD screen) Truly an ongoing usage

PSNR Study Our best quality level was a coasting point rendition of Reinhard administrator Using steady weights to build the Gaussian pyramid we get PSNR which are all things considered 3dB lower

Outline of Presentation Future work

Towards a nine-level installed continuous Reinhard administrator Using more scales takes into account better difference, however geometric arrangement in light of forces of-two are no sufficiently longer To utilize more broad bases, must consider: the connection between the base and the extent of the window the precision of count, which identifies with the measure of the collector used to ascertain the rising and falling geometric arrangement

Towards tone mapping of shading HDR pictures shading ought to be a simple augmentation separate luminance from RGB triplet: tone-delineate luminance utilize the mapped luminance to change the RGB preparatory reenactments are promising

Thank You Questions ?