使用 opencv 进行阈值处理后的文本模糊

Question

I am doing some transformations to capture text from image using tesseract OCR, but, doing so, my text after applying some threshold effect is blurry, so I need some assistance here, a little help.

This is my code:

import cv2
import pytesseract as pyt
import numpy as np

pyt.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"

image = cv2.imread('vacunacion.jpg')
gris = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
thresh = cv2.threshold(gris, 90, 255, cv2.THRESH_BINARY_INV)[1]
imagen_detalle = pyt.image_to_string(thresh, lang='eng',config='--psm 6')
print(imagen_detalle)
cv2.imshow('thresh', thresh)
cv2.imwrite('thresh.jpg', thresh)
cv2.waitKey()

My outpout:

2
> PLAN NACIONAL DE VACUNACION
CONTRA COVID-19 Hooia>—
| DOSIS APLICADAS: 191.480
CORTE 4:00 P.M. MARZO - 03 - 2021
SE AMDRES ay
“a 7 women {A GUNIRA (1345)
twas jf Ae d &
paunaanma Goer anda 760 — 2) ©? maaan
CARTABENA (3.457) BOLI (2500) — OESAN 018)
SURES) a ‘3 NORTE DE SANTANDER (25131
CORDOBA (2968) wo 6 SAATANDER (5.936)
ANTIONUIA (24.8563 TS - BOVAGA (5883),
CUNDALAMARCA (12.0251 TY ‘ ARAUCA C8)
SALDAS (23151 EN, f PS yionann 2
RAMLDA C320 wey Pee casa 1227
oan oe ts
Touma 5.080) BK eae 8 WETA (2444)
ALLE DEL cqUGA (20,160) ~=% se GUAINLA 592)
“m8 Se aN +S UAE (3023
‘ante ca o eco yqurtsig2an
puna 2635) ae SO BARUETAISES!
a
rum fe ‘AMAZDAAS (10548)
Oa
Fuente: Ministerio de Salud y Protecciin Saciat - Datos procesados 03 de marzo - 2021

and this is the final image

The number of applied dosis don't appreciate well in the image, because is so blurry, any technique can be applied here?

This is the original Image.

Answer 1

Tesseract can use the gradients around text as part of its detection, so I'd suggest you avoid thresholding where possible, as it removes the gradients (anti-aliasing, as mentioned by fmw42 ) from the image.

Instead here I'd suggest inverting the image after you grayscale it, and then if necessary you can reduce the brightness to make the more grey text a bit blacker, and increase the contrast to make the grey background a bit more white. If you do need to adjust the brightness and/or contrast I'd suggest using cv2.convertScaleAbs to do so efficiently and avoid integer overflow problems.

Answer 2

• my text after applying some threshold effect is blurry

You applied simple-thresholding and did not achieve the desired result. The two missing parts are:

• 1. You should try other pre-processing methods .
• 2. You should try with other page-segmentation-modes

The other approach for thresholding is taking the binary mask and applying some morphological-transformation . Then you need to display each detected text-region, center the image and apply ocr. Using this approach, you can visualize and see where the problem is.

• 1. Binary Mask

  • 

  • As we can see, the world and the background images were removed from the image.

• 2. Dilation

  • 

  • We applied dilation and bitwise_and for detecing the text more accurately.

• 3. Detecting text regions
  • Tesseract may find same-region multiple times. The idea is to find which part of the text is misinterpreted. Here are some examples:

Region	Result
	~ PLAN NACIONAL DE VACUNACION CONTRA COVID-19 Fam>—
	DOSIS APLICADAS: 191.480
	CORTE 4:00 PM MARZO - 03 - 2021
	YPROVIENGIA —» SANTA MARTA (9501 © — LA GUARA 11345)
	BARRANGRILLA CS 997) ATLANTICD3 754) —- ° * —-— MAGDKLENACIS87)
	GARTAGENA [3.457 }BOLIVAR (2.500) —, . CESAR (3.016)

I found the above results (some of them displayed) using english language. The current language is foreign to me. Therefore you need to configure the tesseract for the language

Code:

---

# Load the library
import cv2
import numpy as np
import pytesseract

# Load the image
img = cv2.imread("u1niV.jpg")

# Convert to HSV color-space
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)

# Get binary mask
msk = cv2.inRange(hsv, np.array([0, 0, 181]), np.array([160, 255, 255]))

# Extract features
krn = cv2.getStructuringElement(cv2.MORPH_RECT, (5 ,3))
dlt = cv2.dilate(msk, krn, iterations=5)
res = 255 - cv2.bitwise_and(dlt, msk)

# OCR detection
d = pytesseract.image_to_data(res, output_type=pytesseract.Output.DICT)

# Get ROI part from the detection
n_boxes = len(d['level'])

# For each detected part
for i in range(n_boxes):

    # Get the localized region
    (x, y, w, h) = (d['left'][i], d['top'][i], d['width'][i], d['height'][i])

    # Draw rectangle to the detected region
    cv2.rectangle(img, (x, y), (x+w, y+h), (0, 255, 0), 5)

    # Crop the region
    cropped = res[y:y+h, x:x+w]

    # Center the region
    cropped = cv2.copyMakeBorder(cropped, 50, 50, 50, 50, cv2.BORDER_CONSTANT, value=255)

    # OCR the region
    txt = pytesseract.image_to_string(cropped)
    print(txt)

    # Display
    cv2.imshow("cropped", cropped)
    cv2.waitKey(0)

# Display
cv2.imshow("res", res)
cv2.waitKey(0)