Preliminary Hough transform implementation #14

khilanravani · 2018-05-28T10:45:30Z

We first extract some information from the image - width, height etc. , convert to luma and compute gradients. Then, for each pixel (x,y), we compute rho = xcos (theta) + ysin (theta). After a few more transformations and finding the bins with highest values, we generate the resulting image.

We first extract some information from the image - width, height etc., convert to luma and compute gradients. Then, for each pixel (x,y), we compute rho = xcos(theta) + ysin(theta). After a few more transformations and finding the bins with highest values, we generate the resulting image.

lehins · 2018-05-28T12:15:40Z

Good job @khilanravani, this looks like good progress. Besides the comments on the code itself here is one thing I'd like to see, namely Hough transform in action. Use a sample image and detect some lines in it, for instance, and then add the resulting image to the repo with example code how you generated it. For example you can use the yield.jpg image, that I think would work great for line detection, similar to how it's done in the documentation of this function: https://www.stackage.org/haddock/lts-10.10/hip-1.5.3.0/Graphics-Image-Processing-Binary.html#v:thresholdWith

khilanravani · 2018-05-28T15:09:25Z

Hello sir! Thank you so much for such a prompt response! Sure sir, I will work it out. I just had one doubt. I didn't exactly understand if you want me to write a separate function for detecting lines or you are asking me to use this hough transform and then apply it to detect lines in an image. Thank you so much for your support sir! Sincerely, Khilan.

…

On Mon, May 28, 2018 at 5:45 PM, Alexey Kuleshevich < ***@***.***> wrote: Good job @khilanravani <https://github.com/khilanravani>, this looks like good progress. Besides the comments on the code itself here is one thing I'd like to see, namely Hough transform in action. Use a sample image and detect some lines in it, for instance, and then add the resulting image to the repo with example code how you generated it. For example you can use the yield.jpg image, that I think would work great for line detection, similar to how it's done in the documentation of this function: https://www.stackage.org/haddock/lts-10.10/hip-1.5.3.0/ Graphics-Image-Processing-Binary.html#v:thresholdWith — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#14 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATaf_HT1DHY3twYh8TiDdd6upL-1B9OVks5t2-ptgaJpZM4UQAVI> .

lehins · 2018-05-28T15:52:08Z

I didn't exactly understand if you want me to write a
separate function for detecting lines or you are asking me to use this
hough transform and then apply it to detect lines in an image.

What I mean is adding a good example that can be used in the documentation of the function, that demonstrates not only how to use the function but also the result produced by it. There is no need to add more sample images to the repo, that's why I suggested using yield.jpg file for the mentioned example.

Also what would be nice is to see Hough transform in action. Simply looking at an image produced by hough function isn't particularly interesting, would you agree? But if there is an example of actually using the result of Hough transform that would be awesome, eg. drawing detected lines on the source image would be cool. For example page 31 of this slides: https://www.uio.no/studier/emner/matnat/ifi/INF4300/h09/undervisningsmateriale/hough09.pdf
At the same time it might be a little bit of a complicated task for now and we can defer till later.

khilanravani · 2018-05-28T16:00:47Z

Sir,

Also what would be nice is to see Hough transform in action. Simply looking at an image produced by hough function isn't particularly interesting, would you agree? But if there is an example of actually using the result of Hough transform that would be awesome, eg. drawing detected lines on the source image would be cool. For example page 31 of this slides: https://www.uio.no/studier/emner/matnat/ifi/INF4300/h09/ undervisningsmateriale/hough09.pdf. At the same time it might be a little bit of a complicated task for now and we can defer till later.

Yes sir, I totally agree with you. It can be a good thing for people to actually see the applications of Hough transform. Yes sir, exactly. It might be a bit difficult to implement it right now. I'll try to spend some time for doing it but if you I agree, it will be better if we postpone it to a little later because actually, I'm running a bit behind my decided timeline and 1st evaluations are coming pretty soon 😅. Although, I assure you that I'll definitely work it out sooner or later. Kindly tell me what you think of this. Thank you, Sincerely, Khilan.

…

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#14 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATaf_Kp6N1BcW587lThwUQ9IoXqd0xTjks5t3B0ogaJpZM4UQAVI> .

lehins · 2018-05-28T16:04:26Z

I'll try to spend some time for doing it but if you I agree, it will be better if we postpone it

Yeah, that's fine. Go ahead and try resolving the rest of the comments I made and than I'll I'll guide you on further improvements that can be made.

khilanravani · 2018-05-28T16:05:27Z

Sure sir! Thank you!

…

On Mon, May 28, 2018 at 9:34 PM, Alexey Kuleshevich < ***@***.***> wrote: I'll try to spend some time for doing it but if you I agree, it will be better if we postpone it Yeah, that's fine. Go ahead and try resolving the rest of the comments I made and than I'll I'll guide you on further improvements that can be made. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#14 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATaf_IU5ToTPg4i445bDXWMZ0b7Ia37tks5t3CAKgaJpZM4UQAVI> .

alpmestan · 2018-05-28T16:12:18Z

@lehins Did you forget to send all your review comments?

lehins

@alpmestan I sure did, thank you for pointing it out :)

lehins · 2018-05-28T12:15:57Z

src/Graphics/Image/Processing/Hough.hs

+
+-- | Some helper functions :
+-- | toImageY : Converts an image to Luma Image
+toImageY :: (ToY cs e, IP.Array arr cs e, IP.Array arr Y Double) =>


This function already exists here: https://www.stackage.org/haddock/lts-10.10/hip-1.5.3.0/Graphics-Image-ColorSpace.html#v:toImageY

Sorry sir. Will correct that.

lehins · 2018-05-28T12:17:29Z

src/Graphics/Image/Processing/Hough.hs

+-- bright (or dark, depending on the color of the background field) spot; by applying a suitable 
+-- filter to the results of the transform, it is possible to extract the locations of the lines in the original image.
+--
+-- <<images/frog_rbg.jpg>>


frog_rbg.jpg That's a terrible image to use as an example, have you looked at it :)

Yes sir, I'll change it.

lehins · 2018-05-28T12:23:15Z

src/Graphics/Image/Processing/Hough.hs

+   distMax :: Double -- Compute Maximum distance
+   distMax = (sqrt . fromIntegral $ (heightMax + 1) ^ (2 :: Int) + (widthMax + 1) ^ (2 :: Int)) / 2 
+
+   accBin = runSTArray $   -- Core part of Algo begins here. Working in a safe way with a mutable array.


In order to write faster code, try researching and understanding difference between "Boxed" and "Unboxed" values in Haskell. For example using runSTUArray here instead of runSTArray together with import of Data.Array.Unboxed instead of Data.Array would give you a significant speed up.
You don't need to do that here now, since in the end there will be no dependency on array package anyways, there are better ways to do that.

lehins · 2018-05-28T12:24:24Z

src/Graphics/Image/Processing/Hough.hs

+   hImage = makeImage (thetaSz, distSz) hTransform
+
+
+test :: IO ()


Test functions like that have no place being in a library module.

Sorry sir, that was for my testing. Forgot to remove it. Will do.

lehins · 2018-05-28T12:27:36Z

src/Graphics/Image/Processing/Hough.hs

+
+   maxAcc = F.maximum accBin  
+   hTransform (x, y) =
+        let l = 255 - truncate ((accBin ! (x, y)) / maxAcc * 255) -- pixel generating function


Pixels with Double precision are kept in the range [0, 1], not [0, 255],which is the range for Word8. So this is certainly a bug.

@khilanravani, I apologize, you original code was correct, hTransform :: (Int, Int) -> Image arr cs Word8, not a Double, so it does need to be 255 - truncate ((accBin ! (x, y)) / maxAcc * 255)

lehins · 2018-05-28T12:32:56Z

src/Graphics/Image/Processing/Hough.hs

+   heightMax = ((cols image) - 1)
+   yCtr = (heightMax `div` 2)
+
+   luma :: Image arr Y Double


There is no point of converting an RGB image into luma. If that is what the function needs than hough function should accept and return Y type images.

Yes sir, actually earlier I used RGBA images but then switched to rgb. I'll change the input type.

What I meant by the comment above is that if hough works with luma images, then it should accept luma images not RGB or RGBA, it needs Y. So you don't need to use toImageY, the user of the hough function will. Here is what a proposed type signature might look like:

hough :: ( MArray arr Y Double, IP.Array arr Y Double ) => Image arr Y Double -> Int -> Int -> Image arr Y Double

Unless I am mistaken, you don't need forall arr a, nor do you need either one of RankNTypes or ScopedTypeVariables extension up top.

I was mistaken, I see now, you do need forall arr and ScopedTypeVariables, but not the RankNTypes

So the signature will be something along those lines:

hough :: forall arr . ( MArray arr Y Double, IP.Array arr Y Double, IP.Array arr Y Word8) => Image arr Y Double -> Int -> Int -> Image arr Y Word8

Ok sir, I'll change it accordingly.

Ok sir, I'll change that.

lehins · 2018-05-28T12:37:34Z

src/Graphics/Image/Processing/Hough.hs

+     , MArray arr Y Double, IP.Array arr Y Double
+     , IP.Array arr RGB Double
+     )
+  => Image arr RGB a


My understanding was that Hough transform works on Binary images, but I haven't really studied it in depth to really understand it. So, convince me that it works correctly and than I'll guide you through on how to improve the code quality and speed.

Sir, as per my understanding, Hough transform is most commonly used in combination with canny, sobel or sometimes with de-noising filters. In that case, the input image is generally pre-processed and converted to binary or grayscale (more preferred). The basic concept of Hough doesn't limit its usage to binary images only although it depends on what exactly we use it for. I checked some other implementations and found that they also use grayscale images in most cases. So, I guess taking rgba as input, converting to luma and applying the algorithm shouldn't be a problem. Please correct me if I'm misunderstanding some point here.

khilanravani

Please ignore this commit.

lehins · 2018-06-03T19:04:07Z

src/Graphics/Image/Processing/Hough.hs

+-- filter to the results of the transform, it is possible to extract the locations of the lines in the original image.
+--
+-- <<images/yield.jpg>>
+--


Could you please include the result image of applying the hough transform to yield.jpg here.
This way a user will know what sort of result to expect.

Sure sir, will do.

Sure sir, will do so.

khilanravani added 2 commits May 28, 2018 15:49

remove outdated file

08c298f

lehins reviewed May 28, 2018

View reviewed changes

khilanravani added 2 commits May 29, 2018 14:21

PR Corrections 1

9b016b3

PR Corrections 1

c2cdcb5

khilanravani commented May 29, 2018

View reviewed changes

lehins reviewed Jun 3, 2018

View reviewed changes

PR Corrections 2

e6b0015

lehins merged commit ea50a9b into lehins:master Jun 12, 2018

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Preliminary Hough transform implementation #14

Preliminary Hough transform implementation #14

khilanravani commented May 28, 2018

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

alpmestan commented May 28, 2018

lehins left a comment

lehins May 28, 2018

khilanravani May 29, 2018

lehins May 28, 2018

khilanravani May 29, 2018

lehins May 28, 2018

khilanravani May 29, 2018

lehins May 28, 2018

khilanravani May 29, 2018

lehins May 28, 2018

lehins Jun 3, 2018

lehins May 28, 2018

khilanravani May 29, 2018

lehins Jun 3, 2018

lehins Jun 3, 2018

khilanravani Jun 4, 2018

khilanravani Jun 5, 2018

lehins May 28, 2018

khilanravani May 29, 2018

khilanravani left a comment

lehins Jun 3, 2018

khilanravani Jun 4, 2018

khilanravani Jun 5, 2018

		hImage = makeImage (thetaSz, distSz) hTransform


		test :: IO ()

Preliminary Hough transform implementation #14

Preliminary Hough transform implementation #14

Conversation

khilanravani commented May 28, 2018

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

lehins commented May 28, 2018

khilanravani commented May 28, 2018 via email

alpmestan commented May 28, 2018

lehins left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

khilanravani left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment