How to improve the code for getting text from an image?

Question

Python

How to improve the code for getting text from an image?
I have this code

from PIL import Image
import pytesseract
import cv2
import os

pytesseract.pytesseract.tesseract_cmd = r'C:\Program Files (x86)\Tesseract-OCR\tesseract.exe'

base_dir = os.path.dirname(os.path.abspath(__file__))
image = base_dir + r'\tmp\test.PNG'
d = Image.open(image)
preprocess = "thresh"

# загрузить образ и преобразовать его в оттенки серого
image = cv2.imread(image)
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

# проверьте, следует ли применять пороговое значение для предварительной обработки изображения

if preprocess == "thresh":
    gray = cv2.threshold(gray, 0, 255,
        cv2.THRESH_BINARY | cv2.THRESH_OTSU)[1]

# если нужно медианное размытие, чтобы удалить шум
elif preprocess == "blur":
    gray = cv2.medianBlur(gray, 3)

# сохраним временную картинку в оттенках серого, чтобы можно было применить к ней OCR
filename_dir = base_dir +"\gray\{}.png".format(os.getpid())
cv2.imwrite(filename_dir, gray)

# загрузка изображения в виде объекта image Pillow, применение OCR, а затем удаление временного файла
text = pytesseract.image_to_string(Image.open(filename_dir))
print(text)
os.remove(filename_dir)

# показать выходные изображения
cv2.imshow("Image", image)
cv2.imshow("Output", gray)

I want it to better recognize text from an image
, here is an example of




gray



Text

fright, tine to put the old girl to work.

When you'll step off the Blue Liner onto the island of Cloverton, your new life will begin.

O Bone Dig
23 - 59 (63)

ME ero rpart

toc mary

v fits te arg Saahe any Mn fof
Poth

How can I do it?

Answer 1

I think that's what you wanted?

Of course, not 100% result, but you can play around with the settings

import cv2
import pytesseract

def text(img, size, chan):
    pytesseract.pytesseract.tesseract_cmd = r'Tesseract-OCR\tesseract.exe'
    scale_percent = int(size)# Процент от изначального размера
    image = cv2.imread(img)
    width = int(image.shape[1] * scale_percent / 100)
    height = int(image.shape[0] * scale_percent / 100)
    dim = (width, height)
    resized = cv2.resize(image, dim, interpolation = cv2.INTER_AREA)
    gray = cv2.cvtColor(resized, cv2.COLOR_BGR2GRAY)        #
    ret, threshold_image = cv2.threshold(gray, chan, 150, 1, cv2.THRESH_BINARY)
    text = pytesseract.image_to_string(threshold_image, config='--psm 11')
    # cv2.imshow("123", threshold_image)
    # cv2.waitKey(0)
    return text

text1 = text("1.png", 350, 150)
print(text1,"\n\n")

text2 = text("2.png", 350, 30)
print(text2,"\n\n")

text3 = text("3.png", 350, 160)
print(text3,"\n\n")

Answer 2

Image preprocessing from CV is very important.
There are many different tricks, for example https://stackoverflow.com/questions/39233823/openc... and many others.
Here they write that the quality of recognition depends on the width of the letter in pixels: https://groups.google.com/forum/#!msg/tesseract-oc...
This is about dpi.
3. Teseract can be passed its tesseract parameters, for example:
conf = u"--psm 11"
text = TS.image_to_string(Image.open('1111.jpg'), config=conf)
psm - Page segmentation modes:
0 Orientation only and script detection (OSD).
1 Automatic page segmentation with OSD.
2 Automatic page segmentation but no OSD or OCR.
3 Fully automatic page segmentation, but no OSD. (Default)
4 Assume a single column of variable length text.
5 A single, uniform block of vertically aligned text is assumed.
6 A single unified block of text is assumed.
7 Treat the image as a single text string.
8 Treat the image as a single word.
9 Treat the image as a single word in a circle.
10 Treat the image as a single character.
11 Sparse text. Find as much text as possible in no particular order.
12 Sparse text from OSD.
13 Raw line. Treat an image as a single text string, bypassing Tesseract-specific hacks.
There will never be a perfectly accurate result, only more errors or fewer errors.

Answer 3

Increase text resolution to 150...300 dpi.
The tesseract is a rather stupid system - the more dots per letter, the better.
on your samples, at best, 75dpi is very small.