View Single Post
Old 1st December 2004, 13:09   #4  |  Link
oledoe
Registered User
 
Join Date: Oct 2001
Posts: 15
waveletreducer.txt (continued)

mode4:
Although all the Wavelet deployment values below a è‡ value will be set to 0, in a Wavelet deployment value, inclination of an incoming signal is maintaining a fixed state, and 0 may eliminate the shake of a delicate vector completely, and it may stop leading to the next block continuously by any techniques other than mode1. (Although it becomes a flat and smooth side completely in Wavelet of Haar, since it is influenced by the surrounding Wavelet deployment value when the other Wavelet is being used, a possibility of becoming a perfect flat and smooth side becomes so low that the value of type is high) . Then, it brings close to 0 more, so that the original Wavelet deployment value is close to 0, it divides so that the original Wavelet deployment value is a value far from 0 and it may not be made to decrease not much from the original value, and it is the degeneration method using ŽZ. It is only made hard to be for this mode not to erase a noise and conspicuous. However, it can be said that it is the mode which cannot break down an outline easily. Moreover, the strength of effectiveness can be adjusted with the value of rate. It is not effective at all by 0. It is as an original formula at 100%. Effectiveness becomes strong rapidly more than by it. It restricts to this mode and a maximum is not in the value of rate. Since a è‡ value is not necessarily raised, the influence to the portion below a è‡ value can be strengthened without affecting a portion with the Wavelet deployment value beyond a è‡ value. However, since a block noise may occur if it raises too much not much, please make it moderate.

Below common-matter In order to make the number of times of wavelet deployment unrestricted as a matter of fact, the resolution of an input picture is once changed into the multiplier of 2. Since wavelet deployment is carried out after being changed into 1024*512, if it is the picture of 320*240 and is the picture of 512*256 and 720*480, it is very heavy. Since weight changes considerably also by levelX, Y, and type, please use it in the range which can be equal to use as a matter of fact. Probably, it will be better to stop and solve, if it did not have confidence considerably in a CPU power that type NR(s) [ level ] the picture of 720*480 by 5*5 4. Although it is NR, only luminosity is NR(ed) fundamentally. A reason is easy and is because five more arguments will increase if you are going to make it a color also NR. If there are too many arguments, please write twice by ŸT“©‚µ‚¢'s, changing [ to NR a color, either ] the value of func. However, in almost all cases, an appearance problem becomes that there is nothing only by NR of luminosity. Although it thinks that waveletReducer.dll and waveletReducer_Athlon.dll are enclosed, it is the thing which sets up a compile option by the object for Pentium, or the object for Athlon, and recompiled it, respectively. the direction which the direction which is not attached at all attaches with the object for Pentium series (it corresponds after PentiumIII) and Athlon becomes Athlon series (3D Now -- loading -- indispensable) It has divided into C++ sauce and C sauce, and VisualC++6.0 and the C section are compiled for the C++ part by VectorC 1.3.3SpecialEdition (since there is no assembler ?‘‚_ capability, it depends on VectorC).

Parameter consideration: mode4 is the mode newly established for the purpose of ?A which does not break down an outline. Since it happened to have thought while groping for the method of NR(ing) without breaking down an outline, it mounted. However, since the effect of NR is low, it may be almost meaningless after all. Moreover, if a è‡ value is raised in order to raise an effect, although it is MASHI, an outline will be ruined from other modes, too. It says or fades. On the characteristic of Wavelet, although it thinks that it cannot fade easily rather than other NR(s) even if it raises a è‡ value earnestly, since it is hard to carry out error diffusion, when still fading, it fades. It says that it fades, or is melted and lost. Even if mode3 seldom raises a è‡ value, it is the mode newly established in order to make a noise easy to take. In raising a è‡ value to a barricade barricade earnestly, there is almost no merit. Merely, since it is only heavy, in such a case, please use mode 0, 1, and 2 obediently. However, although I think that it can NR effectively, how is it by animation from which the outline portion and the portion of BETA have parted clearly? Probably, it will be better to depend on mode2 to animation which it is too old, is full of the noises in a screen, and a noise and an effective signal call half-and-half. In mode3, I think that a noise is probably misunderstood with edge and a miserable result is brought. Moreover, isn't it good like a DVD lip to sauce in which a noise is hardly contained primarily at mode1? In mode3, since the low portion of contrast disappears after all, delicate nuance could not remain easily and I think whether there is any problem which was alike also by mode0. Although it may be hard to use mode 0 and 3 properly, the direction of mode3 thinks a little that the NR effect becomes high. However, as written also in the top, since mode3 is the system which deletes the low portion of a contrast difference, leaving the intense portion of a contrast difference, it thinks that the direction of mode0 is suitable to a scene, an on-the-spot photo, etc. to which the whole screen faded. Probably, it will be better to seldom use, since a block noise tends [ quite ] to come out of type0 except mode1 (calculation is high speed). If it is a type value beyond it, although there is no so big difference, it will change [ that how for it to be / the circumference's of an outline / ruined is various and ]. If a type value is large, although roughness will tend to spread comparatively broadly, the roughness itself instead becomes thin. Although roughness will stop at the circumference of an outline if a type value is low, the contrast of roughness becomes high, and even if it is going to take by 2D NR later, it may suffer troubles.

Future view etc. : * Don't obscure an outline, giving NR effectively about the noise reduction effect. Although it is the mission imposed on NR program, it is a pattern still far from the fresh ground. Still, this program in this time is not practical. But there is [ most ] no case where WaveletNR is the most practical. (Only usual 2D NR will be sufficient for business in general.) It is necessary to carry out to Wavelet truly. It is a rare case. However, it will be that it is not impossible, since a great senior called –^ and Type-G exists. Although the present target of waveletReducer is called Avisynth version Type-G, it is in the state which is going to grope for many things and is going to progress in front little by little. Since a certain amount of indicator for not obscuring an outline for the time being has stood, it will heighten the NR effect more, the following target maintaining this outline. It is groping for the way of fusion to usual 2D NR experimentally for the time being. Although it is a stage still far from practical use, I think that the advantage of skillful ‚_Žg‚¦‚Î Wavelet and the advantage of 2D NR can be united. Of course, 3DNR-ization is considered the back. Although it thinks that it becomes mounting with another function, probably, the formation of 3DNR does not have mounting, since it is the good method of increasing the number of effective samples by mistake. Incidentally the relation of two points which adjoin each other on a level 1 is summarized to Wavelet-ization, and the two points are summarized to one point. On a level 2, two points collected on the level 1 are summarized to two-point Wavelet-ization, and it is summarized to one point. (That is, it is set to NR of the range of 4*4) It is the same as the above about two points settled on the level 2 in the level 3. since -- it continues with 16*16 and 32*32 the range of 8*8, and henceforth Since it is such a system, in a scene from which the color is changing every other dot, for example, it is possible to take whole color MURA, without breaking down change of the color in every dot. (threshold of a level 1 is pressed down lowness in this case, and a level 2 and threshold of 3 are raised) It may be difficult to take a merely still really severe noise something. Wavelet extracts the high frequency ingredient more than fixed of the picture currently operated as a feature point, in Wavelet degeneration, is deleting the portion (that is, a Wavelet deployment value's being low) which is not still so high as for frequency among the extracted feature point, and cuts down MURA, with the portion of the feature left. In the multiplex resolution analysis decomposition NR, when a value is less than a fixed range in two points which adjoin each other among Wavelet deployment values, by making the portion flat and smooth, it turns flat and smooth and the delicate singular point generated in a gently-sloping change is assimilated to a surrounding change (that is, when of the same grade in frequency change). Since it is premised on a noise being not much conspicuous by neither of the cases, the conspicuous noise is conversely extracted as a feature and cannot be NR(ed). If a è‡ value is raised so that the conspicuous portion can be NR(ed), it will delete to other significant portions. It is mode2 that it is going to take a noise even if it deletes the information on a significant portion. In mode2, since Threshold*rate% of value is subtracted also to the Wavelet deployment value beyond a Threshold value, the part effective signal ingredient will decrease. However, the noise which cannot be taken can also be taken in mode 0 and 1. In addition, if rate is 0, it will be called mode2=mode0. Moreover, a Threshold value = as for mode 0 and 2, also in infinity, it becomes equivalent as a result (it is a talk in theory, and on the fact, when a Threshold value is 1280 or more, mode 0 and 2 becomes equivalent).

Some proposals of improvement in the speed are also proposed, and they are due to be mounted one by one. Although it is a translation [ parameter / other NR(s) with ˜M‚_‚è ] in AVIUtil, still more, far one is visible to practical use, and it becomes it that it is sometimes likely to be discouraged. This program is still the thing which is done and between does not have, either. Please watch by the lukewarm eye. Notes: Even if you cover to some patents and you make [ it is sufficient and ] it them, please laugh and allow. Though it suffers a certain damage using this program, I need your help by accountability. Bibliography : Joint establishment publication The signal processing and the image-processing Asakura bookstore by wave let Wavelet transform and its application Softbank C Magazine February, 2003 issue


------------------------
oledoe is offline   Reply With Quote