jasonlfunk / ocr-text-extraction Goto Github PK

For whatever reason, cv2.findContours on line 211 returns three arguments instead of the expected 2. From playing with it, I found that by eliminating the first value would solve the problem.
I am not knowledgeable about cv2 so if this is something obvious someone please say so. (MacOS, cv2 '3.2.0').
Here is the fix that worked for me.
result = cv2.findContours(edges.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
contours, hierarchy = result if len(result) == 2 else result[1:3]

in extract_bv(image)
22 ret,f6 = cv2.threshold(f5,15,255,cv2.THRESH_BINARY)
23 mask = np.ones(f5.shape[:2], dtype="uint8") * 255
---> 24 image, contours, hierarchy = cv2.findContours(f6.copy() ,cv2.RETR_LIST ,cv2.CHAIN_APPROX_SIMPLE)
25 for cnt in contours:
26 if cv2.contourArea(cnt) <= 200:

ValueError: not enough values to unpack (expected 3, got 2)

Unable to resolve this error :

TypeError Traceback (most recent call last)
in
207 X,label=shuffle(X,label,random_state=2)
208 model=get_model()
--> 209 fit_model(model,X,label)

in fit_model(model, X, label)
189
190 datagen.fit(X_train)
--> 191 model.fit_generator(datagen.flow(X_train, y_train, batch_size=32, class_mode="categorical",target_size=(64, 64),color_mode="rgb",save_to_dir=r'/Users/Sanjeeth/Python work space/project1/testing/resnet/', save_prefix='fudus', save_format='jpeg'), steps_per_epoch=len(X_train) / 32, epochs=4)
192
193

Is there any option to set a threshold on how we color the image? I have a file with white-on-black and black-on-white text. The script does a phenomenal job detecting the white-on-black text and converting it. However, it also erroneously inverts many of the black-on-white texts. Is that something that could potentially be tuned and fixed by giving us some parameter options?

ocr.zip

This is a great effort and some tweaking it could be very useful for tesseract OCR.

the problem right now is that resulting text seems almost like it's from an withering ancient book. For example "O" looks almost like "U".

the best way I can put it is texts look almost dirty.

what variables or configs can I alter to retain the clean cut text like shown in the answer at

http://stackoverflow.com/questions/11678542/image-processing-for-ocr-with-leptonica-inverse-color-text

How about the word?

1st is processed image.
2nd is original image.

The script makes some funny things with for e.g. an "8":

Hi @jasonlfunk,
It's a great repo which done decent job on most of the images, I observed that for some color images , it will make black bounding box(inside this text is present), may happen due to size of contour , instead of ((img_x * img_y) / 5) this if I set it to fixed number i.e 10,000(250dpi img) above problem were solved but if text height(colour heading) exceed above threshold and it will not make correct thresholding of that portion
Please let me know is there any robust way to solve above issue.
Thanks