Page 3 of 3 FirstFirst 123
Results 41 to 46 of 46
  1. #41
    Geographic you're right, its hard to tell currently how well it works.
    I tried it but I dont find it easy to do, to adjust the images so that the neural one has same color / contrast / lightness
    Maybe someone can help me ?
    this is my best effort to it


    test.png



  2. #42
    Well not entirely sure.. but i begin to think that my current neural network cannt do what i hoped it to do.
    I will think about different ways to do it, for example the network might count similair pixels (use standard diviation or so)
    and based upon that adjust setting for a "simple" blur operation, (neural net trains the blur radius, for a single pix blur).
    currently i let it directly set the lum value but i dont think thats optimal.. well not sure.... but things have to radical change.
    The network does train i believe, but as is currently it is not aware of borders, it can only slightly single pixel resulting that it learns a blur ratio that would be best for all pixels. that might get rid of the pepper noise but it adds slight blur every where I must think this over.
    The good news is that i got a trainable network (reduced training time a lot today) but now i need to give it the right information and the right tools too...



  3. #43
    How does your current neural network look like? I assume you are using a convolutional neural network, with a few layers. Is that correct?



  4. #44
    I'm using a regression neural network with just 3 layers, with multi threaded training.
    Essentially while normal neural networks answer binary or classify regression nets are optimized to return values -1..0..+1 or larger.
    With about 22 inputs surrounding a single pixel, and code to randomly create lots off train data out of images with those 22 facts.
    i't could have been a lot more facts but.. it takes time to code them
    Originally i had the plan to later stack them (some deep networks are based upon that).
    As i dont have the means (hardware) to train deep networks nor the code, my take on it was to split tasks by smaller nets.
    As small neural nets can be very fast. Combining them keeps them small and fast.

    i know deep neural networks though not yet coded them, i got pretty good idea of their working, but its essentially a bigger black box;
    (often made out of stacked multiple smaller nets, which makes the error feedback calculation really complex. And that adds to their training time. The industry is all in on deep nets, Google made tensorflow while Microsoft improved CNTK, the later might be better but.. i didnt study either, its just that I made a few neural nets at work and got amazed by what they can do on industrial machines.
    Based upon that i'm 100% sure a neural network could do this... but well it has to be made first
    Last edited by Razorblade; 22-Jun-17 at 16:15.



  5. #45
    @ razorbade
    I once saw a very simple but extreme good noise filter (i often wondered why it got so little attention).
    They used blocks 3x3 or 5x5 from which they substracted the brightest and the darkest pixel.
    Then they averaged the remaining pixels of lena, you should know Lena
    Many denoising articles are written about her, its this girl :
    lena.jpg

    What i was thinking that maybe your regression network could trained like the filter i called earlier.
    Although it would be easier to do it without a neural network; this could be just a layer in your stacked network maybe ?
    Well i dont know about deep learning but that filter i wish i had the article link, to show you what it can do.
    But there are so many noise articles about her that its not easy to find back a specific article.
    That filther though was special cause so it was so simple but a lot better then many complex filters.
    I once coded it in openCV (but thats years ago, dont know where that code went)



  6. #46
    Well i think i take a small brake, my research into this will not stop, but for the moment i rewind what i've learned so far and will think about next attempts to do this, i still have a few methods in my mind. I've not given up (i rarely give up).
    So maybe 1 week no coding or so.. taking a few steps back and get fresh insights; working at work as a coder and evening cost me a lot energy.

    @Geographic
    Ah Lena.. yeah i've seen lots of articles with her.
    But i've not heard of the denoiser you talk about, but i can imagine something like that might work, if i understand you correctly you could do this in steps and each time repeat the function (with or without?) altering the darkest/lightest spot (what if that spot is in the current center pixel ?

    If anybody else knows what Geographic is talking about feel free to post a link, cause he doesnt seam to have it anymore.
    It might be something. (I thought i had read all articles about lena denoising, so apparently there is one i missed).


    @ others Feel free to post ideas, or code hints.. in relation to neural nets and blender denoising.



Page 3 of 3 FirstFirst 123

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •