[Octave-bug-tracker] [bug #66698] [octave forge] (image) imgradient may output high values

[Liang Tang]

URL:
  <https://savannah.gnu.org/bugs/?66698>

                 Summary: [octave forge] (image) imgradient may output high
values
                   Group: GNU Octave
               Submitter: lt1234
               Submitted: Wed 22 Jan 2025 05:55:28 PM UTC
                Category: Octave Package
                Severity: 3 - Normal
                Priority: 5 - Normal
              Item Group: Matlab Compatibility
                  Status: None
             Assigned to: None
         Originator Name: lt1234
        Originator Email:
             Open/Closed: Open
         Discussion Lock: Any
                 Release: 9.2.0
        Operating System: Microsoft Windows
           Fixed Release: None
         Planned Release: None


    _______________________________________________________

Follow-up Comments:


-------------------------------------------------------
Date: Wed 22 Jan 2025 05:55:28 PM UTC By: Liang Tang <lt1234>
Hi, I observed a difference between Octave and Matlab in imgradient.  Thanks.


Octave Image imgradient.m may produce high gradient values at the image outer
perimeter.  The attachment mat file contain the image associated with the
numbers listed below.  Other images can be used to demonstrate my observation
if the image does not have a uniform black background.  

grad=imgradient(bw);grad(1:10,1:10)

   667.5148   630.0286   630.0286   626.0288   626.0288   630.0286   630.0286
 627.0072   627.0072   628.0000
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
   3.1623     5.8310     4.0000
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
   1.4142     7.6158     4.0000
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
        0     8.0000          0
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
   1.4142     7.6158     4.0000
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
   4.4721     7.2111     8.0000
   632.0000     8.0000     8.0000     8.0000     8.0000     8.0000     8.0000
   8.2462     6.0000     8.0000
   631.0198    10.7703     8.2462     9.4868     9.4868     8.2462    10.7703
  11.6619     6.3246     7.0711
   629.0199    13.9284     8.9443    10.7703    10.7703     8.9443    13.9284
  10.2956     2.0000     2.0000
   628.0000    13.0384     8.2462     9.0554     9.0554     8.2462    13.0384
   6.3246     7.2111     7.6158
   
Matlab solution:    

    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
4.0000    4.0000         0
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
3.1623    5.8310    4.0000
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
1.4142    7.6158    4.0000
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
0    8.0000         0
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
1.4142    7.6158    4.0000
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
4.4721    7.2111    8.0000
    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000    8.0000
8.2462    6.0000    8.0000
   11.4018   10.7703    8.2462    9.4868    9.4868    8.2462   10.7703
11.6619    6.3246    7.0711
   13.0384   13.9284    8.9443   10.7703   10.7703    8.9443   13.9284
10.2956    2.0000    2.0000
   12.0000   13.0384    8.2462    9.0554    9.0554    8.2462   13.0384
6.3246    7.2111    7.6158

The rest is FYI only.  

imgradientxy is called from imgradient to calculate the gradient.
imgradientxy name is common in Octave/Matlab.  

Octave Image imgradientxy uses conv2    to apply filters.   
Matlab Image imgradientxy uses imfilter to apply filters. 

imfilter has padding options to control what the calculation of gradient at
the image outer perimeter is based on.  
I assume that conv2 uses zero pad.  

The example below shows where I see the differences may be from.  
 
% Octave Image package default
[gradX] = conv2 (double(bw),   fspecial('sobel')',  'same');
gradX(1:10,1:10) 
% imfilter can produce the same as conv2 
[gradX] = imfilter (double(bw), fspecial('sobel')', 'conv');
gradX(1:10,1:10)
 
% But in Matlab imgradientxy, the padding is controlled.  

[gradX] = imfilter (double(bw), fspecial('sobel')', 'replicate',
'corr');gradX(1:10,1:10)  % matlab default 
[gradX] = imfilter (double(bw), fspecial('sobel')', 'replicate',
'conv');gradX(1:10,1:10)  % conv is not an issue.  The padding is.  







    _______________________________________________________
File Attachments:


-------------------------------------------------------
Name: bw.zip  Size: 147KiB
<https://file.savannah.gnu.org/file/bw.zip?file_id=56802>

    AGPL NOTICE

These attachments are served by Savane. You can download the corresponding
source code of Savane at
https://savannah.gnu.org/source/savane-31c1c407b236753eb42459c749b2de08c0e97606.tar.gz

    _______________________________________________________

Reply to this item at:

  <https://savannah.gnu.org/bugs/?66698>

_______________________________________________
Message sent via Savannah
https://savannah.gnu.org/

signature.asc
Description: PGP signature