---

demand

    A need at work , You need to extract the license plate number in the picture . Pictured , The license plate is in the fixed position of the picture . Start using pytesseract, Especially unfriendly to Chinese recognition , After all, it's a foreigner's thing . Colleagues recommend PaddleOCR, The things developed by Chinese people are different , The recognition accuracy can reach 90% above . But both have common problems , White characters on black background cannot be recognized / The range is too small, the identification is not allowed to wait .

---

One . Cut the license plate number area

# np.fromfil  Construct an array from data in a text or binary file 
# cv2.imdecode() Convert the read data ( decode ) In image format ; It is mainly used to recover images from network transmission data 
# cv2.IMREAD_UNCHANGED： Read the full picture , Include alpha passageway , You can write directly -1
img = cv2.imdecode(np.fromfile(imgSrc, dtype=np.uint8), cv2.IMREAD_UNCHANGED)
cropImg = img[y1:y2, x1:x2] # The order is up and down   about 

（ A small picture cut out of the license plate ）

Normally, the small figure cut out from the above is used for OCR, You can recognize the text , result ....

    So Baidu found Zhang Tu , It happens to be black on a white background Writing in the middle , The results magically identify the results 100% The identification is correct . After comparison , I concluded that the white font was unrecognizable , Then invert the small graph .

Two . Process the small picture of the license plate

height, width, deep = cropImg.shape
 gray = cv2.cvtColor(cropImg, cv2.COLOR_BGR2GRAY) # cv2.COLOR_BGR2GRAY  take BGR Convert format to grayscale image 
 dst = np.zeros((height, width, 1), np.uint8) # Generate a pure black picture 
 
 for i in range(0, height):  #  Reverse phase   Turn to black on white 
     for j in range(0, width):
         grayPixel = gray[i, j]
         dst[i, j] = 255 - grayPixel
  # After this step , It's done   Turn to black on white , But the white low background is not the brightest 
  # Reuse cv2.threshold To binarize , Make the black part darker , White is whiter 
ret, img = cv2.threshold(dst, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

    The result of inverting the small graph is as follows , Look carefully and find that the white background is not very white , Reuse cv.threshold To binarize （ Neither black nor white ）, By comparing binary images , The sense of hierarchy comes out .

（ After inversion ）

（ After binarization ）

    Identify again , The recognition result finally came out , But I found the last one 7 Identified as 2, In principle, such simple words should not be recognized wrong . So I thought that the boundary of the picture used to test and identify is very wide , Writing in the middle , Then the small graph is filled with white background boundary 150 Pixel .

3、 ... and . Fill in the boundary

# cv2.BORDER_CONSTANT  Fixed value filling method 
imgsrc = cv2.copyMakeBorder(img, 150, 150, 150, 150, cv2.BORDER_CONSTANT, value=[255, 255, 255])

（ After filling ）

---

Four . Identification steps

ocr = PaddleOCR(use_angle_cls=True, use_gpu=False)  #  Use CPU Preloading , no need GPU
text = ocr.ocr(img, cls=True)
result = str(text[0][1][0]).replace(' license plate ：', '').upper()

（ The final recognition effect ）