Pytesseract not giving expected output

This method uses contour detection to find the area of the number plate, and does a perspective transform on it. It then uses adaptive thresholding to detect the digits and does a medianBlur to get rid of noise that messes up the pytesseract operation.

The blue box is from the original image. The red box is from my number plate recognition haar cascade. The green is the contour detection of the number plate.

Here's the output of the perspective transform. I used the imutils module to do this.

This is the output of the adaptive thresholding and blurring, which I used the skimage module for.

Using this I got the output of:

\fUP1ADN7120

And removing all characters that aren't uppercase or a digit gives:

UP1ADN7120

Which is only one character wrong.

This method isn't too bad, but you could get even better using another method. If this didn't work for you, then you could create a CNN.

Here's the code:

import cv2 import numpy as np import pytesseract import imutils.perspective from skimage.filters import threshold_local plate_cascade = cv2.CascadeClassifier('/usr/local/lib/python3.5/dist-packages/cv2/data/haarcascade_russian_plate_number.xml') img = cv2.imread('GNA0d.jpg', cv2.IMREAD_COLOR) gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) gray = np.array(gray, dtype='uint8') plates = plate_cascade.detectMultiScale(gray, 1.3, 5) for (x,y,w,h) in plates: cv2.rectangle(img, (x,y), (x+w,y+h),(0,0,255),2) roiGray = gray[y:y+h, x:x+w] roiImg = img[y:y+h, x:x+w] blur = cv2.GaussianBlur(roiGray, (5, 5), 0) edges = cv2.Canny(blur, 75, 200) contours = cv2.findContours(edges.copy(), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)[1] contours = sorted(contours, key = cv2.contourArea, reverse = True)[:5] #sort the contours, only getting the biggest to improve speed for contour in contours: peri = cv2.arcLength(contour, True) approx = cv2.approxPolyDP(contour, 0.02 * peri, True) #find contours with 4 edges if len(approx) == 4: screenCnt = approx break orig = roiImg.copy() cv2.drawContours(roiImg, [screenCnt], -1, (0, 255, 0), 2) #do a perspective transform warped = imutils.perspective.four_point_transform(orig, screenCnt.reshape(4, 2)) graywarp = imutils.perspective.four_point_transform(roiGray, screenCnt.reshape(4, 2)) #threshold using adaptive thresholding T = threshold_local(graywarp, 11, offset = 10, method = "gaussian") graywarp = (graywarp > T).astype("uint8") * 255 #do a median blur to remove noise graywarp = cv2.medianBlur(graywarp, 3) text = pytesseract.image_to_string(graywarp) print(text) print("".join([c for c in text if c.isupper() or c.isdigit()])) cv2.imshow("warped", warped) cv2.imshow("graywarp", graywarp) cv2.imshow('img', img) cv2.waitKey(0) cv2.destroyAllWindows() 

Quite a bit was taken from pyimagesearch, who explains it better than I ever could.