OJODIGITAL |
|
|
|
| perfectRAW/perfectBLEND Foro para tratar todo lo relacionado con estos dos programas basados en DCRaw para el revelado de imágenes RAW y el blending de imágenes para aumentar su rango dinámico |
![]() |
|
||||||||||
|
WPG: nuevo algoritmo de interpolación
Abro este hilo para discutir cuestiones relacionadas con WPG, que hasta ahora se venía haciendo en este otro hilo y los mensajes cruzados complicaban su lectura. Si consideráis que está mejor en otro subforo movedlo de sitio.
Copio un extracto a modo de introducción: Cita:
![]() _________________________________ Cita:
![]() Cita:
A mí AHD no me parece un algoritmo de propósito general, de hecho no lo uso nunca. Supongo que será cuestión de gusto, pero para mí sus fallos no compensan nunca sus aciertos. Cita:
![]() El código de Coffin es difícil de entender, pero yo creo que tiene añadidos sobre la versión original. Cita:
Las bandas oblícuas de color que saca VNG RGGB intuyo que no son un problema del algoritmo utilizado, sino de la propia arquitectura de los sensores BAYER o los filtros que lleva delante. Cita:
Cita:
Cita:
Cita:
Cita:
![]() Manuel, no pretendo pasarte el testigo, que ya bastante tienes con perfectRAW. Yo seguiré trabajando en ello, pero en el fondo sé que no podrás resistirte a hacer pruebas y que pondrás aquí los resultados. ![]() ¡Un saludo! Última edición por Lassus; 01-feb-2009 a las 13:45. Razón: Fusión automática de mensajes para prevenir autosubir post |
| Publicidad |
|
||||
|
Lassus, estoy muy liado con el curro y otros temas, pero no pierdo de vista tus resultados con WPG.
Un saludo: |
|
|||
|
Manuel, sigo en ello con el tiempo que saco de aquí y de allá. He mejorado un poco los resultados con unos gradientes más precisos y un umbral adaptivo para hacer que el algoritmo se moje más a la hora de interpolar en una dirección u otra, pero tampoco hay un salto grande con respecto a como estaba.
También he hecho otros dos algoritmos, provisionalmente bautizados como FCD (Fast Conservative Demosaicing) y RGE (Reiterative Gradient Estimation). El objetivo del primero es ser lo más rápido posible (de momento es un 15% más veloz que PPG). Es conservador a propósito para que sea equilibrado y tenga menos falsos colores (el objetivo es que, si se equivoca, no lo haga demasiado para que no se note al 100%). A ISOs altos también va mejor. El resultado es bastante decente, creo. Si te interesa para ponerlo por defecto en perfectRAW dímelo y te lo paso. El segundo se basa en hacer la mejor estimación posible de los gradientes e interpolar en función de ellos sólo en horizontal o en vertical. Creo que éste le va a gustar a Guillermo porque, sin cortar las líneas, tiene menos artefactos que AHD. Es un 60% más rápido que AHD y aún está sin optimizar (tengo una versión "capada" que, obteniendo un resultado muy parecido, tarda la mitad que RGE). A ISOs altos es enfocado, pero crea palitos. Os pongo un pantallazo sobre la marcha de RGE. En esta zona en particular el que lo hace mejor es WPG, y yo diría que FCE, el que tarda menos que PPG, mejora incluso a AHD y AFD: ![]() Muchas siglas juntas, ¿verdad? ![]() Un saludo. |
|
|||
|
Manuel, te respondo por mail.
Para quien lo quiera probar: descarga. Es el dcraw de toda la vida (tutorial aquí) con las siguientes opciones añadidas: -q 4: AFD (el mismo que aparece en perfectRAW). -q 5: FCD (muy rápido, para máquinas viejas o procesar grandes lotes). -q 6: WPG (tiene un pequeño bug, pero sigue siendo bastante válido para todo). -q 7: RGE (para sustituir a AHD en ISOs bajos). Última edición por Lassus; 17-mar-2009 a las 03:09. |
|
||||
|
Muchas gracias, Lassus. En cuanto tenga un rato haré una comparativa completa de tus nuevos algoritmos.
Un saludo: |
|
|||
|
Duro con ellos, que para eso están.
Para ISOs altos iban mejor la primera versión de WPG que te pasé y, por supuesto, AFD. A ISOs bajos distan bastante de ser perfectos, pero sirven para el 98% de las fotos. Por ejemplo en el RAW de la Pentax RGE sigue fallando en las cortinas, pero hace la barandilla perfecta. Los detalles normales (fotos de edicicios, retratos, bodegones) debería hacerlos con bastante solvencia, y los bordes (tanto rectos como curvos) mejor que AHD. |
|
||||
|
Lassus, aver si este finde pruebo esos algoritmos. VCD lo he descartado totalmente, ví unos fallos graves ayer que solo hacía él.
Por ahora uso PPG y AFD. ¿Podrías pasarme el raw de la "botella"? |
|
|||
|
NaVaS, he estado unos días fuera, perdona el retraso.
El archivo para probar ISOs altos es éste. Para probar algoritmos también utilizo éste como muestra de una imagen de mundo real y éste otro para comprobar el comportamiento en casos extremos (la carta central). Ya me comentarás si te resultan útiles los algoritmos que he subido. Aprovecho para exponer mis últimos progresos brevemente: -He escrito un filtro de mediana de radio ajustable y que si se quiere puede discriminar en función de la lejanía del valor respecto de los valores extremos (despeckle) o aplicarse a todos los valores (blur). -He creado un par de nuevos algoritmos, uno de ellos es bastante más rápido incluso que FCD. Del otro ya daré detalles más adelante. -He afinado VCD. Ahora es un algoritmo serio, con gradientes más homogéneos y menos píxeles disparados, y que puede utilizarse perfectamente para todo. Se lo tengo que enviar a Paul. -He optimizado RGE. En verdad la imagen revelada no cambia, pero el proceso es un poco más rápido. Y estoy con: -Un algoritmo especialmente pensado para imágenes ruidosas. -Un nuevo tipo de gradiente para elegir la dirección de interpolación que seguramente exporte a RGE, WPG, etc. Creo que voy a hacer una web con una comparativa en condiciones y donde vaya subiendo las nuevas versiones del dcraw compilado. Mi objetivo es terminar sólo con tres algoritmos: el ultrarrápido, el de propósito general y el de imágenes ruidosas. De los otros también publicaré el código por si alguien lo quiere utilizar pero no los compilaré, que tantas siglas son un verdadero lío. A ver si me pongo este fin de semana, que hasta entonces estoy justito de tiempo. Un saludo. ![]() _______________________________________________ _______________________________________________ He montado los nuevos gradientes y la interpolación del píxel por ratios en RGE y ésta es la diferencia (el archivo proviene de la página de Jacek Góźdź, LinuxPhoto.org): ![]() Dirección más homogénea, menos artefactos, más detalle y más nitidez. ![]() Última edición por Lassus; 07-may-2009 a las 12:03. Razón: Fusión automática de mensajes para prevenir autosubir post |
|
||||
|
Lassus, puedes poner un ejemplo del nuevo RGE con ruido. Tiene una pinta genial.
Un saludo: |
|
|||
|
Manuel, por su diseño el algoritmo hace palitos, aunque parece que menos que antes y desde luego menos que AHD. Mi primera impresión es que da una imagen ligeramente más enfocada que afd, aunque el ruido sea más feo. No es la mejor opción pero al 100% es usable. Aún no pongo el pantallazo porque crea píxeles negros aislados y no sé muy bien por qué, imagino que es cosa de dividir entre los píxeles no captados. Cuando lo arregle la subo.
El gran problema es que trabajar con ratios en coma flotante ralentiza bastante los cálculos. El nuevo RGE es lento, aunque la diferencia creo que merece la pena. Aprovecho y subo un nuevo dcraw compilado ya con versiones de dos de las tres líneas que quiero desarrollar. Este fin de semana tampoco pude, a ver si en el puente saco tiempo, hago la web con la comparativa, le pido permiso a Paul para publicar su VCD modificado y subo los códigos fuente pulidos. Añadidos a dcraw (descarga): -q 4 => Nuevo RGE por ratios. Está sin optimizar en cuanto a velocidad y consumo de recursos, pero creo que ya puede considerarse usable. Debería ser el plato fuerte de todos cuantos algoritmos he hecho, por favor, probadlo y comentad por aquí. -q 5 => Interpolación FDD. Creo que mejora a PPG siendo incluso más rápido. La relación calidad de imagen / tiempo de revelado es muy buena, tardando sólo un 51% más de tiempo que la bilineal (por dar una referencia, en mi ordenador AHD tarda catorce más). -N a b => Filtrado de ruido añadido a propósito para otro hilo. La "a" es el número de pasadas que hace el filtro, mientras que la "b" es el umbral. En principio valores entre 0.25 (más intensidad) y 1 (intensidad normal) son los más adecuados. Cualquier fallo o pregunta comentado por aquí. Un saludo. |
|
||||
|
Comparando por partes. Primero FDD.
Entiendo que el algoritmo FDD de Lassus está diseñado para ser muy rápido y mejorar la calidad de PPG. Veamos si cumple con estos preceptos, en cuyo caso podría pasar a ser el algoritmo por defecto en perfectRAW por méritos propios. Utilizando el ejecutable de Lassus tenemos los siguientes tiempos para el revelado completo:
Ahora vamos a ver si la calidad de la interpolación es igual, mejor o peor que la de PPG. El caso de interpolación más difícil que yo he encontrado hasta ahora es éste: ![]() Como se ve FDD mejora sensiblemente a PPG en las ramitas. Quizás algún detalle lo haga mejor PPG pero en conjunto FDD es mejor. Así que ya tenemos un ganador como nuevo algoritmo de interpolación por defecto en perfectRAW: el FDD de Lassus, ¡enhorabuena! Un saludo: Última edición por ManuelLlorens; 17-mar-2009 a las 22:54. |
|
||||
|
En cuanto a RGE.
Por las pruebas que he hecho (no subiré pantallazos hasta que no arregles los puntos negros, como tú dices) me ha parecido bastante mejor que AHD para imágenes sin ruido (aunque comete el mismo tipo de errores que él cuando falla en la dirección de interpolación y saca peor alguna diagonal) e igual de malo que AHD para imágenes con ruido... bueno quizás un poco mejor que AHD porque sus palitos son más cortos. De momento es bastante lento, pero eso se podrá optimizar. Además el algoritmo de Emil no será más rápido, eso seguro. Lo que sí me ha sorprendido es lo bueno que es FDD comparado con AHD en imágenes sin ruido... sobretodo teniendo en cuenta lo rapidísimo que es. Cuando tenga un rato subo una comparación FDD vs AHD vs RGE en imágenes sin ruido incluyendo los tiempos de procesado. Un saludo: Última edición por ManuelLlorens; 17-mar-2009 a las 23:01. |
|
|||
|
Gracias por los comentarios, Manuel. Lo de las diagonales de RGE me trae de cabeza. Son el motivo de los puntos negros a ISOs altos, así que ahí aún puede afinarse algo. El otro problema que tiene son unos puntitos blancos (quizá los equivalentes en ISOs bajos de los negros) que acompañan a las diagonales por el exterior. Seguiré con ello. Cualquier otro fallo coméntalo por aquí.
FDD está diseñado para interpolar en una sola dirección (horizontal o vertical), como PPG y como AHD, de ahí que los resultados se parezcan. AHD sigue siendo bastante mejor si se mira con lupa, pero para revelado rápido de lotes o en ordenadores lentos FDD va mejor en imágenes ruidosas y en las demás aguanta el tipo. A ver si consigo interpolar la crominancia sin que haga cruces (ahí se gana mucha velocidad) para que no extienda el ruido. Yo ya había hecho unas tablas de tiempo. Hice cinco revelados de cada serie con cada algoritmo y elegí el mejor tiempo en cada archivo (en verdad todos eran muy parecidos). Utilicé un dcraw compilado con DevC++ y el ordenador recién reiniciado. AMD Athlon XP 1700+ con 1024 MB DDR RAM y Windows XP Proffesional SP3: ![]() Como referencia AHD es 13.81 veces más lento que la bilineal. Pentium 4 a 3 GHz con 512 MB DDR RAM y Windows XP Proffesional SP2: ![]() Aquí FDD tarda menos que la bilineal. Tiene que ser cosa de la implementación de Coffin porque desde luego el algoritmo es más complejo. ¡Un saludo! Última edición por Lassus; 18-mar-2009 a las 23:42. |
|
||||
|
Cita:
Un saludo: |
|
|||
|
Subo nueva versión (v.02) del ejecutable. La vieja puede seguir descargándose aquí: v.01.
Cambios:
Así pues: -q 4: RGE -q 5: FDD -q 6: WPG -N a b: Reducción de ruido (a=pasadas, b=umbral). El filtro es el mío, no el de Manuel que se discute en otro hilo. Cita:
Saludos. |
|
|||
|
dear lassus. i wanted to ask how your work on the new demosaicing algorithms is progressing? i have tested all of your implemented algorithms, and i found that RGE gives me the best results with G1 files. i love the look of WPG, but overall there are too many artefacts like zipper noise. RGE manages to decode my files better. lines are rendered nicely. but some artefacts still remain, especially with diagonals. but everything is looking much better than HPHD for example. still regarding smooth lines, DCB from jacek gozdz is unbeatable, at the cost of sharpness and detail of course. i also found that DCB is not the best algorithm for microstructure and natural irregular patterns like foliage. i like RGE more in this case, WPG is even better. the only problem remains with straight diagonal lines, that tend to become staircase artefacts.
it also sems as if highlight recovery didn't work well with RGE yet. i have found some strange artefacts in recovered highlights that were not visible when using AHD. i cannot provide any samples now, i need more time for this, i have had quite a lot to do for work lately. is there any chance that we will see one of your algorithms integrated into perfect raw or somewhere else? i really consider using RGE for all my raw conversion work, but without any GUI only using dcraw as tool it is not very inviting and intuitive. btw. when can we have the first perfect raw 0.7 beta for tryout? ![]() ps. btw i see that you were using one of my old LX3 files for testing. ![]() Última edición por oluv; 29-abr-2009 a las 01:38. |
|
|||
|
Oluv, I'm very sorry for having used one of your files without asking for permission, it's not usual that I don't point the source.
I cannot figure which file you are referring, could you point it so I can add your authorship?I keep working on the algorithms, but I have little time now. Improving RGE is possible (in fact I have already done it). Basically it's a choice between speed and quality, and it's hard to guess which is the optimum combination for both things. I will always choose quality over speed, but probably most people don't want to wait more than five seconds in the demosaicing step. Just out of curiosity, as you said you may use RGE for your RAW conversion work, what would be your choice? I'm aware of the highlight recovery problem but I thought the version I had uploaded had that issue solved. I will check the code again. Thanks for pointing this. As for the straight diagonals I can try to add some kind of support if you like. FDD, RGE and WPG are open to perfectRAW or any other developer I trust. The only limitation is a part of RGE code I don't wish to be published because it's the base for another algorithm. DCB output is too soft for me. It does not create zipper artifacts but produces false colors and deals poorly in noisy images. For example, in the well-known lighthouse photo from the Kodak CD, these are the results: ![]() I will think about a way for combining accurate directions and smooth, artifact free pixel interpolation. Output image will be softer, but it may be prefereable for some cases. Best regards. [BTW, Egon, gracias por darle permiso a Paul para que me enviara el archivo del faro, no sabía que lo habías creado tú o de otra manera te lo hubiera pedido directamente. Y ya que estamos, ¿no tendrás por casualidad más DNGs del Kodak CD o que me puedan servir para tener referencias del PSNR de mi último algoritmo? O si has compilado una alpha de la aplicación para "bayerizar" jpgs, aunque sea sin interfaz y con limitaciones de tamaño, espacio de color, bordes, etc. te estaría eternamente agradecido si me la pasaras]. Última edición por Lassus; 29-abr-2009 a las 20:47. |
|
|||
|
lassus, you missunderstood me, please don't worry because of the raw-file. i originally uploaded it because someone asked me for some LX3 files. this was btw one of my very first test-shots i took with the LX3:
![]() yes i know DCB is quite soft, and i am aware of the false color artefacts, but these are usually very rare in most cased. on the other hand i have lots of files that suffer either from zipper noise or staircase artefacts. jacek claimed that his latest DCB version is not that soft anymore. unfortunately i still haven't got it yet. hopefully i will be able to test it within the next days. i attached an example where i see a definite improvement in RGE over HPHD. but DCB, although very soft, manages to render this scene much more naturally. it also sharpens up quite well, here is the original raw-file if you want to play with it: http://dl.getdropbox.com/u/893528/P1000125.RW2 ![]() of course i am after quality. i don't care much about slowness, i would wait 1 minute if the quality was optimal. hopefully the raw conversion or demosaicing is managable as a batch-process, so waiting shouldn't be an issue. btw, what is RPC? in your example it is nearly indistinguishable from the original. i think it looks even better ![]() what is the other algorithm you are referring to? some time ago in dpreview-forum a guy called "fujicoly" announced a new raw-converter he was working on and asked for some raw-files. in another thread he had showed some results and comparisons which looked very promising, unfortunately i couldn't find them now. i haven't heard anything from him again for several months, so i doubt i ever will... but i am very interested in some high-quality demosaicing and i am willing to pay for it if it really works. ACR is quite crappy with G1 and LX3, C1 is not that bad, but also applies lens-distortion. raw therapee with HPHD is quite ugly with my G1 (besides i have to fight with labyrinth artefacts all the time, but this is another story). so if you are up to something, i am very curious to test it and share the results. |
|
|||
|
Hahaha, I can't believe it!
I was absolutely conviced that the photo you have posted was taken by me!! I even send the RAW as a LX3 file example to fujicoly. Probably I have downloaded it one or two days before buying the LX3 and just put the file among the rest of my own RAWs. The first days I did a lot of test shots and since then I believe it was my photo. Sorry for the misunderstanding! ![]() ![]() ![]() The algorithm I referred is RPC. It is a theoretical approach, intended for reaching the best PSNR measure. It has a nearly-perfection color rendition and deals nice with false colors, but the final output may suffer from severe zipper artifacts and overshooted pixels, so do NOT use it with usual city shots like the last you have posted. It's also veeery slow. I'm currently using it for portraits, landscapes, macro, irregular lines architecture (cathedrals and similar) and any kind of photo in which demosaicing artifacts are visible at 100%. It will be nice to see a new DCB version. Wait a minute and I will compile a new dcraw with it. _______________________________________________ _______________________________________________ Oluv, these are the options of the compiled dcraw: -q 4 remains as RGE v2. -q 5 is a new version of FDD, slightly faster and accurate. -q 6 remains as WPG v2. For testing purposes (for the last three be sure to have a huge amount of free RAM memory available): -q 96: FDD_mod. Created two minutes ago. It's an old FDD version with diagonal gradients. Obviously I haven't had time for testing it yet. It's just an experiment for making comparisons with original FDD so we can decide which approach is better.-q 97: RGE v3. Much slower, worse high ISO behaviour, but better straight line rendition. For me it doesn't pay the difference. It does not include any kind of diagonal support yet. -q 98: JDDFD. An alpha version of the future high ISO algorithm that will replace WPG. There is a big room for improvement at each step as I haven't seriously started with this algorithm "in depth". The only up-to-date step is the refining part. Luminance and chrominance are extracted from old filters. It's just for having an idea of the powerful tool it could be. Usage "-q 98 -J a b c d e", where a is the number of luminance noise removal passes, b is luminance noise removal treshold, c is the number of chominance smoothing passes, d is edge sensing strength and e are the number of an improved EECI refinement passes. The current kernel is a 5x5 R-G, B-G median in the chrominance space, filter to come must be much faster. A good starting point is "-q 98 -J 1 0.75 2 200 1". -q 99: RPC. Described above. There are three options for changing its behaviour: -R a b c, where a is the treshold where the algorithm decides to weight horizontal and vertical lines instead of choosing between them, b are the number of median passes performed to the gradients (should be lower for microstructures, higher for straight lines) and c are the refinement passes (lower leads to worse color rendition and high ISO performance but may avoid overshoted pixels). Default (-R 2.15 2 2 1) is already optimized for most cases. [Edited: "a" option seems to be unoperative, sorry]. I have included a batch encoder for processing a huge amount of files with one single click. It is not mine but from Speek website. Última edición por Lassus; 30-abr-2009 a las 00:28. Razón: Fusión automática de mensajes para prevenir autosubir post |
|
|||
|
Cita:
. this is the natural history museum btw. i have taken hundrets of photos from there, as the place is near my work and seems perfect for camera-testing with lots of micro-structure:Museums Photo Gallery by oluv at pbase.com i also used the museum as motive to discover some assymetrical lens-softness with my G1-lens. the softness was hardly visible with "normal" photos. Cita:
jacek hasn't responded yet. but i assume the new version will be as sharp as the first versions, but with improved zipper-noise. i was able to achieve some nice results with older versions of DCB, like this, only foliage and the irregular stone-texture are looking "odd". but straight lines are decoded nicely: bildercache - professionell kostenlos sofort und schnell Bilder hochladen und bearbeiten, drehen, spiegeln, verkleinern later jacek decided to go for a stronger post processing, that blurred the images too much for my taste. the best would be a kind of "intelligent" algorithm, that automatically manages to switch between regular and irregular textures, and decodes both differently. but this would be hard to achieve i guess. thanks a lot for your effort. i will play with it during weekend, when i have finally a bit more time. Cita:
![]() ps. you also mentioned fujicoly. any idea what happened to him? have you seen his results? he never ever responded again to my emails but i was really expecting a lot from the previews he posted._______________________________________________ _______________________________________________ a quick test, i think i still like RGE most, still haven't tried the latest DCB though. with this G1-example RGE shows least artefacts and has only slightly lower detail rendition than RPC or WPG. HPHD from gabor seems to be somewhere between RPC and WPG regarding artefacts. sometimes even RPC looks better than HPHD. ![]() what is your favorite algorithm? you said that you use the LX3, and from my experience with LX3 images they have a very similar "look" to the G1 with the difference that they are slightly more noisy, about 1.5 - 2 stops. but otherwise they would be hard to distinguish one from each other. haven't played with high-iso files yet. will have a look later and try the new JDDFD. Última edición por oluv; 30-abr-2009 a las 18:17. Razón: Fusión automática de mensajes para prevenir autosubir post |
|
|||
|
Oluv, I use RGE or RPC depending on the image. When I feel lazy (so most of the time) I also use a B/W algorithm that outputs an useable image without further post-processing.
![]() I'm in Paris right now and I don't know if I will read the forum everyday. When I come back on wednesday I will answer your post in detail. ![]() Última edición por Lassus; 01-may-2009 a las 16:26. |
|
|||
|
wow, paris! how nice
![]() i played with some other of my raw files and did another interesting comparison. there are 3 points of interest in this crop. - the roof in the upper half. only RPC and HPHD is able to decode it without false-color moiree. - the umbrella to the right. in both RPC and HPHD it shows artefacts, not sure how you would call them, but this is typical for red colors and HPHD. it seems as if RPC does the same thing with reds. - the fine diagonal structure to the bottom. only RPC manages to renders it, the other algorithms soften it too much. ![]() i mean these are extreme examples. in most cases these differences wouldn't be even noticeable. but what bothers me are the artefacts, that nearly all algorithms produce with "normal" structures like the following, any idea why this is happening? this is the edge bottom of the image, could it be that the algorithms get confused because of CAs, only DCB renders it without problems: ![]() here is the original RAW-file if you want to have a look at it: http://dl.getdropbox.com/u/893528/P1000498.RW2 finally here one full image converted with RGE. i am quite satisfied with it as i think it is looking quite sharp and natural, what do you think? (click for full-size): ![]() |
|
|||
|
Oluv, yes, Paris is nice, but it's about the tenth time I'm here this year... I'm a bit tired!
![]() The fullres image you have posted is fully useable. It seems RGE has done a good job. ![]() I've been thinking about the algorithm you are looking for and I have several ideas that could work. Let me try them on wednesday. CAs are not a problem for RGE and RPC. If you take a look in detail RGE fails when interpolation direction is more or less 45º. I can add diagonal interpolation, but finding the optimal values will take me some time. The problem with RPC is much more complex. I cannot solve that artifacts without changing the whole interpolation method. Simply it does not fit that kind of images. Have you tried it with noisy images or pictures where texture is important? What are FDD_mod (-q 96) results with the branches of the last pictures? |
|
|||
|
interestingly FFD-Mod really does a better job on the branches:
![]() so far i haven't tried it yet. but i played a bit with it now and it is slightly less detailed than RGE for example. it also has some problems with faint nearly horizontal lines, RGE resolves them better. RPC would be indeed my favorite algorithm if detail extraction and texture was my main concern, i like how it renders fine distant foliage that are beyond the resolving frequency of the sensor. here an example (strongly sharpened) that shows how nice RPC looks with very fine texture, RGE starts to look blocky and pattern-like: ![]() but the problem is that as soon as there are lots of straight lines, RPC starts to produce artefacts. a combination of different algorithms or an "intelligent" algorithm would be best, that is able to distinguish between real edges and simple texture, but i have no idea if something like this is possible at all .if i had the time to devolp an image for achieving the perfect look, i would probably convert it with different algorithms and then compose the best parts from each of them together... ![]() Última edición por oluv; 04-may-2009 a las 14:35. |
|
||||
|
I just will say that Emil and Paul have recently done great improvements in their interpolation algorithm and I think we will see some results soon.
BR, |
|
|||
|
manuel, any idea when we will see some results from emil and paul?
so far there is not a single raw-converter that is really able to handle G1-raw files well. i mean if one wants mushy blurred images, he could also use silkypix, but i think a new algorithm included into a new perfect raw which also has adjustable green equilibration would be a big hit. ![]() |
|
|||
|
Oluv, it took me two minutes to create FDD_mod from an old FDD version so do not expect serious results from it.
It was just for showing that adding diagonal gradients can avoid zipper artifacts along diagonals.It shouldn't be difficult to merge RPC for texture and RGE for edges. I will try it this weekend. Jazek has released a new version of DCB. It's much better now but it continues to be soft. Gradient work is impressive (Manuel, try the Pentax file with it). A mix between his gradients and my interpolation method would be nice, I think. I suppose the idea is to add Emil and Paul interpolation to perfectRAW, so you should be able to use it with green channel equilibratiton. From tomorrow I will connect more often. Best regards. |
|
|||
|
DCB is not perfect either. have a look at these crops:
![]() at a particular angle DCB fails to decode edges, you can also see it well in the last crop with the window-arc. funnily all these examples look much better with RGE for example. although you claimed before that 45° gradients are a problem for RGE. it is not always that RGE produces stairsteps with 45° angles. it seems to depend on the colors, on the position in the image and maybe something else. no idea... ![]() i played a bit with the latest DCB, but i think that meanwhile i prefer RGE, because it renders a better texture while still not perfect. but i am very curious to see what you guys will come up with. slightly off-topic: may i ask how you did this famous lighthouse-example from the kodak-cd? is this a raw-file from some particular camera? how do you compare "original" to "converted" then, i mean what is "original"? or is it some tiff, that you put through a virtual bayer-process and demosaic it afterwards? Última edición por oluv; 05-may-2009 a las 16:32. |
|
||||
|
The lighthouse image is a Kodak-CD image. For testing interpolation algorithms it is first "mosaiced" and later "demosaiced" by the interpolation algorithm, so you can compare the original image with de mosaiced/demosaiced one.
Of course, the mosaiced process simulates what a bayer sensor does. The Kodak-CD images are not perfect (they are compressed and oversharpened for my taste) but they are a classic in the literature. The lighthouse image is important becuase it shows how good is the interpolation algorithm reconstructing nyquist textures. BR, Última edición por ManuelLlorens; 05-may-2009 a las 16:44. |
|
||||
|
Cita:
BR, Última edición por ManuelLlorens; 05-may-2009 a las 17:24. |
|
|||
|
here is another comparison that shows the difference well. RGE looks fine on the left side of the arc, but creates stairsteps on the right side. DCB is the other way round, it is smooth on the right, but doesn't manage to render the arc on the left side.
why does the left side look fine with RGE, but the right side not? is it the contrast of the edge? ![]() |
|
|||
|
Cita:
I don't know if Jacek reads this forum, but I guess the problem on those diagonals is the use of bilinear interpolation for estimating the green pixels at positions (row-1,col-1), (row-1,col+1), (row+1,col-1) and (row+1,col+1). A simply weighted estimation will solve the problem. It a curious behaviour. I hope to improve that by adding diagonal interpolation (I'm doing some preliminary tests). I will try to have the finished version this weekend. Best regards. ![]() |
|
|||
|
i already talked to jacek about the diagonal-problem. he said it is a limitation of his demosaicing method. as you mentioned correctly DCB is based on bilinear interpolation.
maybe you could suggest your idea to him. my mathematical knowledge is too low to give him advices. of course i am still interested in seeing your progress with RGE ![]() |
![]() |
| Marcadores |
| Herramientas | |
| Desplegado | |
|
|