How to read the text by easyocr correctly?

Question

I am trying to read images from a camera module and so far I got to process the image this way using adaptive filtering. Besides, I did a lot of manipulation to crop the ROI and read the text. However, it is reading the number but not the units beside the numbers, which are comparatively small in size. How do I solve this problem?

import easyocr 
import cv2
import numpy as np

import matplotlib.pyplot as plt
import time
import urllib.request
url = 'http://192.168.137.108/cam-hi.jpg'
while True:
    img_resp=urllib.request.urlopen(url)
    imgnp=np.array(bytearray(img_resp.read()),dtype=np.uint8)
    image = cv2.imdecode(imgnp,-1)
    image = cv2.medianBlur(image,7)
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)    #to gray convert
    th3 = cv2.adaptiveThreshold(gray_image,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,\
                cv2.THRESH_BINARY,11,2) #adaptive threshold gaussian filter used
    kernel = np.ones((5,5),np.uint8)
    opening = cv2.morphologyEx(th3, cv2.MORPH_OPEN, kernel)
    

    x = 0   #to save the position, width and height for contours(later used)
    y = 0
    w = 0
    h = 0

    cnts = cv2.findContours(opening, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
    cnts = cnts[0] if len(cnts) == 2 else cnts[1]
    threshold =  10
    font = cv2.FONT_HERSHEY_SIMPLEX  
    org = (50, 50) 
    fontScale = 1 
    color = (0, 0, 0)
    thickness = 2
        
    for c in cnts:
        
        approx = cv2.approxPolyDP(c,0.01*cv2.arcLength(c,True),True)
        area = cv2.contourArea(c)   
        if  len(approx) == 4 and area > 100000:   #manual area value used to find ROI for rectangular contours
        
            cv2.drawContours(image,[c], 0, (0,255,0), 3)
            n = approx.ravel()
            font = cv2.FONT_HERSHEY_SIMPLEX
            (x, y, w, h) = cv2.boundingRect(c)
            old_img = opening[y:y+h, x:x+w]  #selecting the ROI
            width, height = old_img.shape
            cropped_img = old_img[50:int(width/2), 0:height] #cropping half of the frame of ROI to just focus on the number
            
            new = reader.readtext(cropped_img)   #reading text using easyocr
            if(new == []): 
                text = 'none'
            else:
                text = new
                print(text)
#                 cv2.rectangle(cropped_img, tuple(text[0][0][0]), tuple(text[0][0][2]), (0, 0, 0), 2)
                if(text[0][2] > 0.5): #checking the confidence level
                    
                    cv2.putText(cropped_img, text[0][1], org, font, fontScale, color, thickness, cv2.LINE_AA)        
            cv2.imshow('frame1',cropped_img)
    key = cv2.waitKey(5) 

    if key == 27:
        break

cv2.waitKey(0)
cv2.destroyAllWindows()

the code in your question is broken. in python, indentation is syntax. please [edit] and fix. — Christoph Rackwitz, Jun 05 '22 at 18:30
Yes. I did that. While copying from python to stackoverflow, indentation got messed up. However, I just need idea how to solve this problem. There is nothing wrong in the code, it's just that easyocr is not able to read certain texts. — Ritika Shrestha, Jun 05 '22 at 20:46
please review [mre]. a screenshot of _output_ is not suitable _input data_ to run your code and reproduce the issue. — Christoph Rackwitz, Jun 05 '22 at 23:21
@RitikaShrestha Apply Otsu threshold to binarize the image and find the largest contour, which hopefully should be the screen. Crop the screen region over the grayscale and enhance it. All the text would be in dark against white background — Jeru Luke, Jun 08 '22 at 10:27
@Ritika Shrestha. You wanted to read 5.32 micro watts or just 5.32? — toyota Supra, Jun 08 '22 at 10:35
5.32 uW. easyocr doesn't read greek but atleast for now, uW would be fine too. @toyotaSupra — Ritika Shrestha, Jun 08 '22 at 10:54
@RitikaShrestha The unit is a non-ASCII character, I doubt it will recognize that. Are the other characters/text recognized well? — Jeru Luke, Jun 08 '22 at 14:06

score 0 · Answer 1 · answered Jun 08 '22 at 18:49

This is the best I could get. The Greek symbol 'mu' is identified as 'p'. I also tried searching for Greek language model related to easyocr but could not find any.

Here is what I did:

Performed Otsu Threshold on the entire image
Selected contour with largest area and cropped it
Converted the cropped image to LAB color space
Manually performed binary threshold on A-channel

I got the following:

Passed this image as input to easyocr:

from easyocr import Reader
reader = Reader(['en'])

# input is the cropped image
results = reader.readtext(crop_img)

# convert to LAB space
lab = cv2.cvtColor(crop_img, cv2.COLOR_BGR2LAB)

# threshold on A-channel
r,th = cv2.threshold(lab[:,:,1],125,255,cv2.THRESH_BINARY_INV)

# create copy of cropped image
crop_img2 = crop_img.copy()

# draw only first 5 results for clarity
# borrowed from: https://pyimagesearch.com/2020/09/14/getting-started-with-easyocr-for-optical-character-recognition/
for (bbox, text, prob) in results[:5]:
  (tl, tr, br, bl) = bbox
  tl = (int(tl[0]), int(tl[1]))
  tr = (int(tr[0]), int(tr[1]))
  br = (int(br[0]), int(br[1]))
  bl = (int(bl[0]), int(bl[1]))
  crop_img2 = cv2.rectangle(crop_img2, tl, br, (0, 0, 255), 3)
  crop_img2 = cv2.putText(crop_img2, text, (tl[0], tl[1] - 20), cv2.FONT_HERSHEY_SIMPLEX, 1.1, (0, 0, 0), 5)

score 0 · Answer 2 · answered Aug 23 '22 at 17:33

If you try to clear the image and pass path to below method it work try

def text_extraction(image, lang_code='en'):
    reader = easyocr.Reader([lang_code], gpu=False)
    roi = cv2.imread(image)#[85:731, 265:1275]
    output = reader.readtext(roi)
    # it returns list of tuple with ([x,y coordinates],text,text_threshold)
    return output

How to read the text by easyocr correctly?

2 Answers2