FEAT: Image Deconvolution Functions #1881

9prady9 · 2017-07-21T16:17:42Z

Iterative deconvolution algorithms

Landweber
Richardson-Lucy

Linear deconvolution algorithms
~~- [ ] Wiener~~

Tikhonov

Wiener has been disabled for now. I haven't had a break finding the bug in that. We can enable the API in future.

pavanky · 2017-07-21T16:25:42Z

src/api/c/deconvolution.cpp

+    auto ONES   = createValueArray(dims, scalar<T>(1));
+    auto HALF   = createValueArray(dims, scalar<T>(0.5));
+    auto midVal = arithOp<T, af_mul_t>(mxArr, HALF, dims);      // B/2
+    auto diff   = arithOp<T, af_sub_t>(prev, midVal, dims);     // |I(n-1) - B/2|


You are not calling abs anywhere.

missed that. ~~Need to look into why tests passed, or if the abs is needed or not~~. Thanks. I know why they passed - this(jansson) algorithm is not being tested right now as I don't have reference data. ITK has Landweber and I added tests only for that as of now.

New changes pushed.

umar456 · 2018-04-08T15:30:26Z

docs/details/image.dox

+
+Iterative Deconvolution Algorithms
+
+The following table shows the iteration update equations of the respective deconvolution algorithms.


Can you make this more descriptive? It doesn't talk about what the algorithm does.

Sure. I think few pictures can help illustrate the case too.

umar456 · 2018-04-08T15:31:12Z

docs/details/image.dox

+    - \f$ \alpha \f$ is the relaxation factor
+    - \f$ \otimes \f$ indicates the convolution operator
+
+\note The type of output \ref af::array from deconvolution will be double if the input array type is


Format lines to be 80 characters per line?

umar456 · 2018-04-08T15:32:05Z

examples/image_processing/CMakeLists.txt

@@ -13,6 +13,7 @@ project(ArrayFire-Example-Image-Processing

 find_package(ArrayFire)

+set(ASSETS_DIR "${ArrayFire_SOURCE_DIR}/assets")


You probably need to update to reflect our most recent changes to the example folder.

umar456 · 2018-04-08T15:33:28Z

include/af/image.h

+  C++ Interface for Iterative deconvolution algorithm
+
+  \param[in] in is the blurred input image
+  \param[in] ker is the kernel(point spread function) known to have caused the blur in the system


80 characters per line.

umar456 · 2018-04-08T15:34:33Z

include/af/image.h

+/**
+  C++ Interface for Iterative deconvolution algorithm
+
+  \param[in] in is the blurred input image


What type of images do you expect? Are there any restrictions?

Addressed. Check out the docs for this function to see the updates.

umar456 · 2018-04-08T15:51:09Z

include/af/image.h

+
+  \ingroup image_func_iterative_deconv
+ */
+AFAPI array iterativeDeconv(const array& in, const array& ker,


I feel like the name should be deconvIterative and deconvInverse. It emphasizes the operation you are performing.

Having deconv as prefix sounds odd and reverse. Having Deconv as a suffix has an ease to the name when we want to talk about the function or refer to it somewhere. Also, it has a nature flow as to how the name translates to the brief description of the algorithm.

iterativeDeconv - Iterative Deconvolution
inverseDeconv - Inverse Deconvolution

sounds better compared to

deconvIterative - Iterative Deconvolution
deconvInverse - Inverse Deconvolution

umar456 · 2018-04-08T15:54:00Z

src/api/c/deconvolution.cpp

+// cuFFT/clFFT works with any data size, but is optimized
+// for size decomposition with prime factors up to
+// 7.
+const dim_t GREATEST_PRIME_FACTOR = 7;


Does this mean that the OpenCL and the CPU versions will behave differently? We should avoid doing that if that is the case.

There is no difference in behavior except that I am using prime factor that each backend is optimized for when we use ffts. The error metric computed against the test data is same across all backends. The prime factor is used to determine the padding length, nothing else.

umar456 · 2018-04-08T16:16:31Z

test/inverse_deconv.cpp

+        std::vector<OutType> goldData(nElems);
+        ASSERT_EQ(AF_SUCCESS, af_get_data_ptr((void*)goldData.data(), goldArray));
+
+        ASSERT_EQ(true, compareArraysRMSD(nElems, goldData.data(), outData.data(), 0.03));


How did you get the 0.03 threshold for this operation? Wouldn't this be depended on the type of the input and output data and its range?

I guess what I am asking is that have you done any comparison on the image itself and looked at the results at the pixel level and calculated why there are differences. Sometimes there are subtle bugs which cannot be viewed or appear when you look at metrics like RMSD

It would depend on the input range unless we bring them into the same range, which I am doing both while generating the test data from ITK and after calculating the output in ArrayFire tests.

I have checked the images in GIMP and I didn't notice any odd/huge artifacts. If there is any difference, it is probably some(random spread out) pixels here and there I think.

umar456 · 2018-04-08T16:29:33Z

src/backend/cpu/kernel/pad_array_borders.hpp

+namespace kernel
+{
+static dim_t
+calcIdx4Border(const dim_t i, const dim_t lb,


This function should really be optimized. This is being used a lot in your function and it is using some expensive operations (lots of branching and modulus). You should think more carefully about how to implement this.

umar456 · 2018-04-08T16:43:16Z

docs/details/image.dox

+    - \f$ \gamma \f$ is a user defined noise variance constant
+
+
+\note The type of output \ref af::array from deconvolution will be double if the input array type is


What is the range of the float values. Is it 0-1 or 0-255?

The algorithm/function itself is niether dependent on the range of the data nor alters the range of the data.

Addressed feedback

9prady9 · 2018-06-06T10:49:25Z

@umar456 I have cleaned the style a bit further. It is ready for merge unless there is more feedback.

The following iterative algorithms have been added * Landweber - Tests PASS * RichardsonLucy - Tests PASS The following linear inverse algorithms have been added * Tikhonov - Tests PASS Wiener Filter - Tests FAIL - Disabled temporarily, The function is also not exposed at the moment. An example demonstrating both iterative and inverse deconvolution has been added as well. Unit tests using grayscale images have also been added.

umar456 · 2018-04-19T19:51:21Z

test/CMakeLists.txt

@@ -273,3 +273,5 @@ make_test(SRC where.cpp)
 make_test(SRC wrap.cpp)
 make_test(SRC write.cpp)
 make_test(SRC ycbcr_rgb.cpp)
+make_test(SRC inverse_deconv.cpp)


Combine the tests into one file.

Will in another PR while doing cleanup with other functions.

umar456 · 2018-04-19T20:07:13Z

src/api/c/deconvolution.cpp

+            richardsonLucy(currentEstimate, paddedIn, P, Pc,
+                           iters, normFactor, odims);
+            break;
+        default:


This should throw an error if the wrong algorithm value is sent.

9prady9 added the feature label Jul 21, 2017

9prady9 added this to the v3.6.0 milestone Jul 21, 2017

pavanky reviewed Jul 21, 2017

View reviewed changes

9prady9 force-pushed the deconvolutions branch from 28da69f to 85e5f49 Compare July 21, 2017 16:43

9prady9 changed the title ~~[WIP] FEAT: Deconvolution Functions~~ [WIP] FEAT: Image Deconvolution Functions Jul 21, 2017

9prady9 force-pushed the deconvolutions branch 5 times, most recently from 1c29b6c to a0274c1 Compare July 27, 2017 09:35

9prady9 force-pushed the deconvolutions branch from a0274c1 to d4809c3 Compare August 18, 2017 14:12

pavanky changed the base branch from devel to master August 23, 2017 15:21

9prady9 force-pushed the deconvolutions branch from d4809c3 to 7442516 Compare August 23, 2017 16:01

9prady9 force-pushed the deconvolutions branch 2 times, most recently from 4cc2163 to 5c25132 Compare September 12, 2017 11:40

9prady9 force-pushed the deconvolutions branch from 5c25132 to 4b7d5e1 Compare October 20, 2017 15:53

9prady9 force-pushed the deconvolutions branch 2 times, most recently from d1eec82 to 94a167c Compare December 13, 2017 03:37

9prady9 force-pushed the deconvolutions branch from 94a167c to e844c5d Compare December 28, 2017 07:40

9prady9 changed the title ~~[WIP] FEAT: Image Deconvolution Functions~~ FEAT: Image Deconvolution Functions Dec 28, 2017

9prady9 force-pushed the deconvolutions branch from e844c5d to e0f8044 Compare January 11, 2018 17:36

9prady9 force-pushed the deconvolutions branch 2 times, most recently from 815b1cc to d851935 Compare January 25, 2018 05:10

9prady9 force-pushed the deconvolutions branch 2 times, most recently from 468351c to 0f04d60 Compare February 13, 2018 08:18

9prady9 requested review from umar456 and pavanky February 13, 2018 09:53

9prady9 force-pushed the deconvolutions branch from 0f04d60 to 7fa4d2c Compare February 20, 2018 03:04

9prady9 added the ready to merge label Feb 20, 2018

9prady9 force-pushed the deconvolutions branch from 7fa4d2c to 630c58c Compare February 20, 2018 03:41

9prady9 force-pushed the deconvolutions branch from cfefbdc to 0e9f9ef Compare February 22, 2018 14:29

9prady9 requested a review from umar456 February 22, 2018 14:29

umar456 requested a review from syurkevi February 23, 2018 00:54

9prady9 force-pushed the deconvolutions branch from 0e9f9ef to 5d1ba38 Compare February 23, 2018 03:58

umar456 removed this from the v3.6.0 milestone Feb 27, 2018

9prady9 force-pushed the deconvolutions branch from 5d1ba38 to a6e185a Compare March 8, 2018 04:58

9prady9 force-pushed the deconvolutions branch from a6e185a to 10d1df0 Compare April 3, 2018 04:02

9prady9 removed the request for review from syurkevi April 3, 2018 04:03

umar456 previously requested changes Apr 8, 2018

View reviewed changes

9prady9 force-pushed the deconvolutions branch from 10d1df0 to 956936c Compare April 9, 2018 06:36

9prady9 added in-progress and removed ready to merge labels Apr 12, 2018

9prady9 force-pushed the deconvolutions branch 2 times, most recently from ab90678 to 64cfac8 Compare April 19, 2018 11:35

9prady9 requested a review from umar456 April 19, 2018 11:36

9prady9 removed the in-progress label Apr 19, 2018

9prady9 force-pushed the deconvolutions branch from 64cfac8 to fb0be21 Compare May 3, 2018 13:01

9prady9 force-pushed the deconvolutions branch from fb0be21 to 0a6332a Compare June 6, 2018 10:48

9prady9 force-pushed the deconvolutions branch from 0a6332a to 77c64b2 Compare June 15, 2018 11:06

9prady9 force-pushed the deconvolutions branch from 77c64b2 to 16ba10a Compare July 11, 2018 04:30

9prady9 mentioned this pull request Jul 11, 2018

Remove buffer alloc calls from sparse. Pass params by value #2235

Merged

9prady9 added the ready to merge label Jul 13, 2018

umar456 reviewed Jul 13, 2018

View reviewed changes

umar456 approved these changes Jul 13, 2018

View reviewed changes

9prady9 merged commit d88f626 into arrayfire:master Jul 13, 2018

9prady9 deleted the deconvolutions branch July 13, 2018 07:00


		Iterative Deconvolution Algorithms

		The following table shows the iteration update equations of the respective deconvolution algorithms.

		@@ -13,6 +13,7 @@ project(ArrayFire-Example-Image-Processing

		find_package(ArrayFire)

		set(ASSETS_DIR "${ArrayFire_SOURCE_DIR}/assets")

		- \f$ \gamma \f$ is a user defined noise variance constant


		\note The type of output \ref af::array from deconvolution will be double if the input array type is

FEAT: Image Deconvolution Functions #1881

FEAT: Image Deconvolution Functions #1881

Uh oh!

Conversation

9prady9 commented Jul 21, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 Jul 21, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 Apr 9, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 Apr 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 Apr 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 Apr 9, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

9prady9 commented Jun 6, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

9prady9 commented Jul 21, 2017 •

edited

Loading

9prady9 Jul 21, 2017 •

edited

Loading

9prady9 Apr 9, 2018 •

edited

Loading

9prady9 Apr 18, 2018 •

edited

Loading

9prady9 Apr 18, 2018 •

edited

Loading

9prady9 Apr 9, 2018 •

edited

Loading