Recognize text from picture

JonB

@sodoku For things like enumerations, you can usually check the swift version of docs, which tells you the value. Otherwise, you can often look up source code.

For minimumTextHeight, both swift and ObjC say this is a float. The fact that it is readwrite/nonatomic/assign is not important.
So, usually this would just be
req.minimumTextHeight = 32.5
or whatever you want...

It is often helpful to explore objects in the console, since this can tell you what you're working with. For instance, if you type req. in the console, you will see autocomplete of all known attributes. Usually you need to treat objc properties as function calls -- so to check minimumTextHeight, you'd use req.minimumTextHeight(). But to set, you can treat the property as a python attribute and assign directly. In some cases, you may need to use the set_propertyName_(value) convention.

Where things get tricky is where the declared type is another object (in which case you have to provide the right type of object), or a structure. Structures can be tricky because objc_util often screws up the type encodings, and you have to manually override. Structures get turned into python STRUCTUREs, and you access fields normally like you would with a python object (no () needed).

Re question 2:
Per the docs, the results of a request will be VNRecognizedTextObservation objects. This is a subclass of VNRectangleObservation.

@interface VNRecognizedTextObservation : VNRectangleObservation <-- colon here means inherits from

If you look up VNRectangleObservation, you will see it has the following attributes
bottomLeft
bottomRight
topLeft
topRight
Which are declared as CGPoint, which is a structure that has an .x and .y fields.

for result in req.results():
    x = result.bottomLeft().x
    y = result.bottomLeft().y
    w = result.topRight().x-x
    h = result.topRight().y-y
    print('({},{},{},{}) {}'.format(x,y,w,h, result.text())

You could draw the image into an image context, and then also stroke a rectangle.. something like this...(not tried).

with ui.ImageContext(ui_image.size()) as ctx:
   ui_image.draw()
   for result in req.results():
      vertecies = [(p.x, p.y) 
                               for p in [result.bottomLeft()
                                        result.TopLeft()
                                        result.TopRight()
                                        result.BottomRight()
                                        result.bottomLeft()]
      pth = ui.Path.moveTo(*vertecies[0]) %initial point
      for p in vertecies[1:]:
         pth.line_to(*p)  
      ui.set_color('red')
      pth.stroke()
      x,y = vertecies[0]
      w,h =(vertecies[2].x-x), (vertecies[2].y-y)
      ui.draw_string(result.text(), rect=(x,y,w,h), font=('<system>', 12), color='red')
   marked_img = ctx.get_image()
   marked_img.show()

JonB

I realized that result will also have a .boundingBox() attribute which would make some of this a little simpler.
That is a CGrect, consisting of .origin (in turn consisting of .x and .y and .size containing .w and .h.
In that case you could use ui.Path.rect.

JonB

Okay, my previous reply was full of errors... here is a working version, which adds red boxes around each result, along with the text

language_preference = ['fi','en','se']

import photos, ui, dialogs
import io
from objc_util import *

load_framework('Vision')
VNRecognizeTextRequest = ObjCClass('VNRecognizeTextRequest')
VNImageRequestHandler = ObjCClass('VNImageRequestHandler')

ACCURATE=0
FAST=1

def pil2ui(pil_image):
    buffer = io.BytesIO()
    pil_image.save(buffer, format='PNG')
    return ui.Image.from_data(buffer.getvalue())

selection = dialogs.alert('Get pic', button1='Camera', button2='Photos')

ui_image = None

if selection == 1:
    pil_image = photos.capture_image()
    if pil_image is not None:
        ui_image = pil2ui(pil_image)
elif selection == 2:
    ui_image = photos.pick_asset().get_ui_image()

if ui_image is not None:
    print('Recognizing...\n')

    req = VNRecognizeTextRequest.alloc().init().autorelease()
    req.recognitionLevel= ACCURATE# accurate
    req.setRecognitionLanguages_(language_preference)
    handler = VNImageRequestHandler.alloc().initWithData_options_(ui_image.to_png(), None).autorelease()

    success = handler.performRequests_error_([req], None)
    if success:
        for result in req.results():
            print(result.text())
    else:
        print('Problem recognizing anything')

with ui.ImageContext(*tuple(ui_image.size) ) as ctx:
   ui_image.draw()
   for result in req.results():
      cgpts=[   result.bottomLeft(),
                result.topLeft(),
                result.topRight(),
                result.bottomRight(),
                result.bottomLeft()  ] 
      vertecies = [(p.x*ui_image.size.w, (1-p.y)*ui_image.size.h) for p in cgpts]
      pth = ui.Path()
      pth.move_to(*vertecies[0]) 
      for p in vertecies[1:]:
         pth.line_to(*p)  
      ui.set_color('red')
      pth.stroke()
      x,y = vertecies[0]
      w,h =(vertecies[2][0]-x), (vertecies[2][1]-y)
      ui.draw_string(str(result.text()), rect=(x,y,w,h), font=('<system>', 12), color='red')
   marked_img = ctx.get_image()
   marked_img.show()

mikael

@JonB and @sodoku, just a note that I tried a different route, where I first used a rectangle recognizer to isolate the numbers, and only then used text recognition.. The results were not impressive, but I can try to find the code for reference, if you think it might be useful.

JonB

Here is another solution... I use a rectangle detection and a perspective correction to crop the puzzle. This gives much better detection, though not perfect. The recognition is pretty good, though it has troubles with 1’s on their own.... turn into Ts of all things. Some additional work in the clean function might fix common problems.

I’m using images from https://github.com/prajwalkr/SnapSudoku/tree/master/train

I suspect doing some CIFiltering first will probably improve things.

from objc_util import *
import ui

VNImagePointForNormalizedPoint=c.VNImagePointForNormalizedPoint
VNImagePointForNormalizedPoint.argTypes=[CGPoint, c_int, c_int]
VNImagePointForNormalizedPoint.restype=CGPoint


ui_image=ui.Image.named('image2.jpg')
ui_image.show()

CIImage=ObjCClass('CIImage')
ci_image=CIImage.imageWithCGImage_(ui_image.objc_instance.CGImage())

CIPerspectiveCorrection=ObjCClass('CIPerspectiveCorrection')
f=CIPerspectiveCorrection.perspectiveCorrectionFilter()
f.inputImage=ci_image
o=f.outputImage()

load_framework('Vision')
VNRecognizeTextRequest = ObjCClass('VNRecognizeTextRequest')
VNDetectRectanglesRequest = ObjCClass('VNDetectRectanglesRequest')
VNImageRequestHandler = ObjCClass('VNImageRequestHandler')

req=VNDetectRectanglesRequest.alloc().init().autorelease()
req.maximumObservations=2
req.minimumSize=0.5
req.minimumAspectRatio=0.7
req.quadratureTolerance=30

handler = VNImageRequestHandler.alloc().initWithData_options_(ui_image.to_png(), None).autorelease()

success = handler.performRequests_error_([req], None)
try:
   result=req.results()[0]
   nm=lambda p :VNImagePointForNormalizedPoint(p,int(ui_image.size.w),int(ui_image.size.h))
   f.topLeft = nm(result.topLeft())
   f.topRight = nm(result.topRight())
   f.bottomLeft = nm(result.bottomLeft())
   f.bottomRight = nm(result.bottomRight())
   o=f.outputImage()

   with ui.ImageContext(o.extent().size.width, o.extent().size.height) as ctx:
     UIImage.imageWithCIImage_(o).drawAtPoint_( CGPoint(0,0))
     ui_image2=ctx.get_image()
   ui_image2.show()
except:
   print('bounding rec not found...results wont work')
   ui_image2=ui_image
'''now, detect rectangles again...'''
handler = VNImageRequestHandler.alloc().initWithData_options_(ui_image2.to_png(), None).autorelease()
req0 = VNRecognizeTextRequest.alloc().init().autorelease()
req0.recognitionLevel= 0# accurate
req0.usesLanguageCorrection=True
req0.customWords=[str(a) for a in range(10)]

#req0.maximumObservations=81
#req0.minimumSize=.1
success = handler.performRequests_error_([req0], None)
with ui.ImageContext(*tuple(ui_image2.size) ) as ctx:
   ui_image2.draw()
   for result in req0.results():
      cgpts=[result.bottomLeft(),
                                        result.topLeft(),
                                        result.topRight(),
                                        result.bottomRight(),
                                        result.bottomLeft()] 
      vertecies = [(p.x*ui_image2.size.w, (1-p.y)*ui_image2.size.h) for p in cgpts]
      pth = ui.Path()
      pth.move_to(*vertecies[0]) 
      for p in vertecies[1:]:
         pth.line_to(*p)  
      ui.set_color('red')
      pth.stroke()
      x,y = vertecies[0]
      w,h =(vertecies[2][0]-x), (vertecies[2][1]-y)

      ui.draw_string(str(result.text()), rect=(x,y,w,h), font=('<system>', 12), color='red')
   marked_img = ctx.get_image()
   marked_img.show()
   
def bbcenter(bb):
   return((9*(bb.origin.x+bb.size.width/2)-0.5), 
          (9*(bb.origin.y+bb.size.height/2)-0.5) )
def clean(results):
   cleaned=[]
   for r in results:
      col,row=bbcenter(r.boundingBox())
      approx_num_ch=(r.boundingBox().size.width*9)
      txt=str(r.text()).replace(' ','')
      if approx_num_ch<=1:
         if len(txt) == 1:
             cleaned.append(((round(col),round(row)),txt))
         else:
             cleaned.append(((round(col),round(row)),'-1'))
      else: #more than one char
         col-=(len(txt)-1)/2
         col=round(col)
         row=round(row)
         for ch in txt:
           if ch in [str(a) for a in range(10)]:
             cleaned.append(((col,row),ch))
           else:
             cleaned.append(((col,row),'-1'))
           col+=1
   return cleaned


import numpy as np
puzzle=np.zeros([9,9])
for c,v in clean(req0.results()):
   puzzle[c]=int(v)
print(np.flipud(puzzle.T))

mikael

@JonB, thanks, very nice. I have noted and wondered about how difficult number 1 is to recognize... Not very exotic, is it? But in my experiments it looked like the simple heuristic of ”if the result is something else than 1-9, assume it is a 1” would work pretty well for Sudoku.

mikael

@JonB, can you open up this one a little bit?

approx_num_ch=(r.boundingBox().size.width*9)

JonB

The *9 is because if the initial rectangle detection and crop works, then each square is approx 1/9 width. So the approx number of squares a rectangle covers tells us how many characters it should have... I was getting many cases where 1 got read as Te, or some other two character value, even though the width was less than one box... so I wanted to have special handling for narrow boxes, as that is probably a 1, while wide boxes could have multiple characters because the bounding box legitimately spans adjacent boxes.

mikael

@JonB, there’s something in the math here I do not quite get. I would expect something like:

num_char = r.bbox/(full_bbox/9) = r.bbox * 9 / full_bbox

Thus looks like you are missing the division?

JonB

The results of vision are always provided as normalized coordinates — meaning the full box is always 1.
For drawing, you have to then multiply by image width/height.

Since the perspective correction both fixes perspective and crops — 1/9 is the size, roughly, of a single cell.

mikael

@JonB, now I understand, thank you.

sodoku

I have a quick question in regards to the original ocr post how do I print the text as one single csv list, I tried but I have been getting a list of lists instead of one single list

This is a snippet of the code example I think needs to be altered

  success = handler.performRequests_error_([req], None)
    if success:
        for result in req.results():
            print(result.text())
    else:
        print('Problem recognizing anything')

JonB

results=[str(result.text()) for result in req.results()]

print(results)

Or maybe

print(','.join(results))

ccc

There is a lot of good code here... It would be really awesome if there was a GitHub repo to stitch it all together into an app.

mikael

@ccc, do you mean this one? PRs are always welcome.

chrillek

Hi,
I'm aware that this thread is about a year old. But maybe someone can nevertheless alighten me. I'm trying to do a similar thing in JavaScript for Automation (JXA), and I see this line in your example:

for result in req.results():
            print(result.text())

translated to JavaScript, that's

results.forEach(r => {
      console.log(r.text);
 })

and that works like a charm. I'm just wondering why, since according to Apple's documentation, the results object doesn't even have a text property, only string (cf. https://developer.apple.com/documentation/vision/vnrecognizedtext?language=objc)

I was first wondering if text is perhaps a nice Python thing, but since the same works in JavaScript, I'm sure that I'm missing something obvious in Apple's documentation. Does anyone know what (and where I should be looking)?

Thanks a lot in advance
Christian

JonB

Does string not work?

Often there are undocumented or decrecates features available in objc objects. Often we just poke around using autocomplete (which ultimately uses some of the introspection objc features of the objc runtime (which let you get a list of methods or instance vars, etc)

chrillek

Does string not work?

It does, but only in a very convoluted way, like so:

results.forEach(r => {
      console.log(r.topCandidates(1).js[0].string.js)
})

The js in the middle is required to convert the ObjC array returned by topCandidates to a JavaScript array (and again to convert the NSString returned by string to a JS string). But using string directly at r does not work.

we just poke around using autocomplete

I gues that happens in XCode (the poking around)?

JonB

No, the exploration happens in pythonista, in the console. Once you have an object, dir(variable) lists the methods and such, or frankly just typing a letter and autocomplete suggestions does it's thing.

If you're not using a bridging library like
https://github.com/TooTallNate/NodObjC
I'd suggest that you do, since it might take care of a lot of the annoying bits like converting every type to js equivalents, and let's you access some of the dynamic introspection stuff that makes objc pretty neat.

Under the hood, there are objc runtime functions that let you get lists of method names. For instance,see
https://github.com/jsbain/objc_hacks/blob/master/print_objc.py
For how you can do it in python. Or, look at the NodObjC code for class.js and core.js -- it looks like it does something similar, using the objc copy_methodsList, etc, and adds those as J's callable functions to the prototype. Then your favorite J's debugger ought to show you what is there...

JonB

In this case, looking at the headers for VNTextObservation shows the text attribute.

https://github.com/xybp888/iOS-Header/blob/master/13.0/Frameworks/Vision.framework/VNRecognizedTextObservation.h