Tesseract-据称容易将图像转换为错误的数字

Question

Please find below some images that tesseract recognized it incorrectly.

47 is recognized as "4]".

55 is recognized as "S55".

90 is recognized as "IQ".

I thought the images are pretty good and should be easy to be recognized by Tesseract. But the results turn out to be wrong. The code I used is shown below.

import cv2
import pytesseract
from PIL import Image
import glob

for i in glob.glob('*.png'):
    img = cv2.imread(i, 0)
    tessdata_dir_config = '--tessdata-dir "C:\Program Files (x86)\Tesseract-OCR\" --psm 10'
    result = pytesseract.image_to_string(Image.fromarray(img), config=tessdata_dir_config)
    print result

Does anyone know what is going on and how to improve the performance?

Answer 1

Okay, I find an answer for my question. It seems that Tesseract doesn't like bold characters, so you have to erode the black part of the characters a little bit. But beware of that cv2.erode will erode white part of the characters, so we have to use cv2.dilate to achieve the objective.

for i in ['47-4].png', '55-S55.png', '90-IQ.png']:
    img = cv2.imread(i, 0)

    ### After apply dilation using 3X3 kernal. The recognition results are improved.##
    kernel = np.ones((3, 3), np.uint8)
    img = cv2.dilate(img, kernel, iterations=2)

    cv2.imwrite("./output/" + i[:-4]+'_dilate.png', img)
    tessdata_dir_config = '--tessdata-dir "D:\Program Files\Tesseract-ocr\" --psm 10'
    result = pytesseract.image_to_string(Image.fromarray(img), config=tessdata_dir_config)
    print result

I would like to see if there are any better analysis to this question. So I would let it open for a while and choose the best answer.

Answer 2

I had the problem of reading text from android device screens. On some devices it worked on others didn't. I found in tesseract documentation that it has something to do with image dpi.

Tesseract works best on images which have a DPI of at least 300 dpi, so it may be beneficial to resize images. For more information see the FAQ.

So I used resize function of cv2 to rescale the image.

    path = "/home/share/workspace/NNW4JJ4T4LR4G66H_ZTE_Blade_L5/clock_present_cropped.png"
    path2 = "/home/share/workspace/NNW4JJ4T4LR4G66H_ZTE_Blade_L5/clock_present_cropped_2.png"
    crop_img2 = cv2.imread(str(path))
    img_scaled = cv2.resize(crop_img2, None, fx=0.5, fy=0.5, interpolation=cv2.INTER_LINEAR)
    cv2.imwrite(str(path2), img_scaled)
    crop_img2 = Image.open(path2)
    result = pytesseract.image_to_string(crop_img2)

Now it works well with all devices.