CODEWITHSUNDEEP
pythonopencvcontour
ant0nisk
ant0nisk

Reputation: 591

drawContours fails to create a correct mask in OpenCV 3.1.0

I want to split the characters of an Image using OpenCV in order to train a Tesseract model.

I am using version 3.1.0 (because of a Macports upgrade - meh..), and the documentation (for Python) is still not very clear/well-documented.

Here is what I do:

  1. Binarize the Image in order to be ready to find the contours
  2. Find the contours (which works - at least I get non-zero results)

  3. For each contour:

    1. Create a mask (this probably fails - gets zeros)
    2. Extract the part from the original image using the mask (should work but the mask fails, so this doesn't work yet)

The new version of OpenCV has somewhat different syntax as well, so this makes it even more tricky sometimes.

Here is my code:

def characterSplit(img):
    """
    Splits the characters in an image using contours, ready to be labelled and saved for training with Tesseract
    """

    # Apply Thresholding to binarize Image
    img = cv2.GaussianBlur(img, (3,3), 0)
    img = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 75, 10)
    img = cv2.bitwise_not(img)

    # Find Contours
    contours = cv2.findContours(img, cv2.RETR_EXTERNAL , cv2.CHAIN_APPROX_TC89_KCOS, offset=(0,0))[1]

    # Iterate through the contours
    for c in xrange(len(contours)):
        mask = numpy.zeros(img.size)

        cv2.drawContours(mask, contours, c, (0, 255, 0), cv2.FILLED)    # Mask is zeros - It might fail here!

        # Where the result will be stored
        res = numpy.zeros(mask.size)

        # Make a Boolean-type numpy array for the Mask
        amsk = mask != 0

        # I use this to copy a part of the image using the generated mask.
        # The result is zeros because the mask is also zeros
        numpy.copyto(res, img.flatten(), where = amsk)

        ## (... Reshape, crop and save the result ...)

As far as I know, the mask should be of the same size as the original image. But should it also have the same shape? For instance, my image is 640x74 but the way I create my mask matrix, my mask is 1x47360. Maybe this is why it fails... (but doesn't throw any errors)

Any help is appreciated!

Upvotes: 0

Views: 1062

Answers (1)

ant0nisk
ant0nisk

Reputation: 591

I ended up doing what Miki proposed in the comments. I used cv::connectedComponents to do the character splitting. Here is the corresponding code, for anyone who is interested:

def characterSplit(img, outputFolder=''):
    # Splits the Image (OpenCV Object) into distinct characters and exports it in images withing the specified folder.

    # Blurring the image with Gaussian before thresholding is important
    img = cv2.GaussianBlur(img, (3,3), 0)
    img = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 75, 10)
    img = cv2.bitwise_not(img)

    output = cv2.connectedComponentsWithStats(img, 8, cv2.CV_16S)
    n_labels =  output[0]
    labels = output[1]
    stats = output[2]

    for c in xrange(n_labels):
        # Mask is a boolean-type numpy array with True in the corresponding region
        mask = labels == c

        res = numpy.zeros(img.shape)
        numpy.copyto(res, img, where=mask)

        # The rectangle that bounds the region is stored in:
        # stats[c][0:4] -> [x, y, w, h]

        cv2.imwrite("region_{}.jpg".format(c), res)

Hope this helps!

Upvotes: 1

Related Questions