sikuli-driver team mailing list archive

[Nash <695616@xxxxxxxxxxxxxxxxxx>]

Public bug reported:

I have noticed inconsistency in the Sikuli X text recognition and
matching, especially when text on the screen/region are integers. Below
are the steps to reproduce the problem scenario:

Step1: In your web browser open 'TextOnTheScreen.png' from the following
web-link (http://qoydkw.bay.livefilestore.com
/y1pSr0oDju9ndvMzJJfo_qWD2qEvMdDDwoKquMNWWkUYAxgrsVcXmETY8yW_LqIieOKXgwlTt8yxe8eQDY-
SCySCFljo59mSSIC/TextOnTheScreen.png?psid=1)

Step2: To the below sikuli script, provide region coordinates of the 'TextOnTheScreen.png' image which is on your web browser. Note! region coordinates must be significantly bigger because you will zoom-in & zoom-out of the visible content on your browser in-order to carry-out multiple text recognition tests.
--------------------------------------------------------
r = Region(“TextOnTheScreen”)
t1 = "534438"
t2 = "534177"
t3 = "4438"
t = t2
m = r.exists(t,0)
if m:
if m.text() == t:
 popup("Exact match")
else:
 popup(m.nearby().text())
else:
popup("Nothing found at all!")
--------------------------------------------------------

Step3(TestCase1): Now, run the above sikuli script with t = t2. You might notice one or more of the following:
- Integers such as 7 & 8 are incorrectly recognised
- All the integers-as-text are recognised correctly
- Only few integers are recognised correctly

Questions: 
- Does the text recognition depend on how much OCR gets trained? 
- Does the font-type & font-size matter? 
- Should there be a mechanism for validating an exact-match for text recognition?

Step4(TestCase2): Now, back to your browser where the
'TextOnTheScreen.png' image is visible. Try to zoom-out (ctrl&-) the
image and run the above sikuli script with t = t2. You should now notice
variation in the text recognition results.

Step5(TestCase3): Repeat 'TestCase2' with different zoom-levels in your
browser, you might get varied text recognisation results.

Step6(TestCase4): Now, run the above sikuli script with t = t3. You
might notice that integers-as-text are not recognised at all or with
luck it might recognise some integers-as-text?

Hope this provide some test-cases for troubleshooting inconsistency in
text/integer recognition.

Questions:
- Should text and integer recognition be independent operations? Should there be a choice for the application developer if she/he wants to combine and/or isolate such operations where-seen-fit? Would such separation lead to reduced-OCR-complexity/accuracy/better-control over the OCR?
- How would one reliability obtain an exact string match for e.g. with click("String") operation?

** Affects: sikuli
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Sikuli
Drivers, which is subscribed to Sikuli.
https://bugs.launchpad.net/bugs/695616

Title:
  Sikuli X 1.0rc1 - Inconsistency in text recognition and matching, especially with integers-as-text!

Status in Sikuli:
  New

Bug description:
  I have noticed inconsistency in the Sikuli X text recognition and matching, especially when text on the screen/region are integers. Below are the steps to reproduce the problem scenario:

Step1: In your web browser open 'TextOnTheScreen.png' from the following web-link (http://qoydkw.bay.livefilestore.com/y1pSr0oDju9ndvMzJJfo_qWD2qEvMdDDwoKquMNWWkUYAxgrsVcXmETY8yW_LqIieOKXgwlTt8yxe8eQDY-SCySCFljo59mSSIC/TextOnTheScreen.png?psid=1)

Step2: To the below sikuli script, provide region coordinates of the 'TextOnTheScreen.png' image which is on your web browser. Note! region coordinates must be significantly bigger because you will zoom-in & zoom-out of the visible content on your browser in-order to carry-out multiple text recognition tests.
--------------------------------------------------------
r = Region(“TextOnTheScreen”)
t1 = "534438"
t2 = "534177"
t3 = "4438"
t = t2
m = r.exists(t,0)
if m:
if m.text() == t:
 popup("Exact match")
else:
 popup(m.nearby().text())
else:
popup("Nothing found at all!")
--------------------------------------------------------

Step3(TestCase1): Now, run the above sikuli script with t = t2. You might notice one or more of the following:
- Integers such as 7 & 8 are incorrectly recognised
- All the integers-as-text are recognised correctly
- Only few integers are recognised correctly

Questions: 
- Does the text recognition depend on how much OCR gets trained? 
- Does the font-type & font-size matter? 
- Should there be a mechanism for validating an exact-match for text recognition?

Step4(TestCase2): Now, back to your browser where the 'TextOnTheScreen.png' image is visible. Try to zoom-out (ctrl&-) the image and run the above sikuli script with t = t2. You should now notice variation in the text recognition results.

Step5(TestCase3): Repeat 'TestCase2' with different zoom-levels in your browser, you might get varied text recognisation results.

Step6(TestCase4): Now, run the above sikuli script with t = t3. You might notice that integers-as-text are not recognised at all or with luck it might recognise some integers-as-text?

Hope this provide some test-cases for troubleshooting inconsistency in text/integer recognition.

Questions:
- Should text and integer recognition be independent operations? Should there be a choice for the application developer if she/he wants to combine and/or isolate such operations where-seen-fit? Would such separation lead to reduced-OCR-complexity/accuracy/better-control over the OCR?
- How would one reliability obtain an exact string match for e.g. with click("String") operation?