Detect and Recognize Car License Plate from a video in real time

Recognizing Car License Plate is a very important task for a camera surveillance-based security system. We can extract the license plate from an image using some computer vision techniques and then we can use Optical Character Recognition to recognize the license number. Here I will guide you through the whole procedure of this task.

Requirements:

opencv-python 3.4.2
numpy 1.17.2
skimage 0.16.2
tensorflow 1.15.0
imutils 0.5.3

Example:

Input:

Output:
29A33185

Approach:

Find all the contours in the image.
Find the bounding rectangle of every contour.
Compare and validate the sides ratio and area of every bounding rectangle with an average license plate.
Apply image segmentation in the image inside validated contour to find characters in it.
Recognize characters using an OCR.

Methodology:

To reduce the noise we need to blur the input Image with Gaussian Blur then convert the it to grayscale.
Find vertical edges in the image.
To reveal the plate we have to binarize the image. For this apply Otsu’s Thresholding on the vertical edge image. In other thresholding methods we have to choose a threshold value to binarize the image but Otsu’s Thresholding determines the value automatically.
Apply Closing Morphological Transformation on thresholded image. Closing is useful to fill small black regions between white regions in a thresholded image. It reveals the rectangular white box of license plate.

Above 4 steps are performed by preprocess method of class PlateFinder

def preprocess(self, input_img): 
  
    imgBlurred = cv2.GaussianBlur(input_img, (7, 7), 0) 
      
    # convert to gray 
    gray = cv2.cvtColor(imgBlurred, 
                        cv2.COLOR_BGR2GRAY)  
      
    # sobelX to get the vertical edges 
    sobelx = cv2.Sobel(gray, cv2.CV_8U,  
                       1, 0, ksize = 3)   
      
     # otsu's thresholding 
    ret2, threshold_img = cv2.threshold(sobelx,0, 255, 
                           cv2.THRESH_BINARY + cv2.THRESH_OTSU) 
 
    element = self.element_structure 
    morph_n_thresholded_img = threshold_img.copy() 
    cv2.morphologyEx(src = threshold_img, 
                     op = cv2.MORPH_CLOSE, 
                     kernel = element,  
                     dst = morph_n_thresholded_img) 
      
    return morph_n_thresholded_img 

To detect the plate we need to find contours in the image. It is important to binarize and morph the image before finding contours so that it can find more relevant and less number of contours in the image. If you draw all the extracted contours on original image, it would look like this:

This step is performed by extract_contours method of class PlateFinder

def extract_contours(self, after_preprocess): 
  
    _, contours, _ = cv2.findContours(after_preprocess, 
                                      mode = cv2.RETR_EXTERNAL, 
                                      method = cv2.CHAIN_APPROX_NONE) 
    return contours 

Now find the minimum area rectangle enclosed by each of the contour and validate their side ratios and area. We have defined the minimum and maximum area of the plate as 4500 and 30000 respectively.

Code: Methods validating the area and side ratios of minimum area rectangle are validateRatio and
preRatioCheck of class PlateFinder:

def validateRatio(self, rect): 
  
    (x, y), (width, height), rect_angle = rect 
 
    if (width > height): 
        angle = -rect_angle 
          
    else: 
        angle = 90 + rect_angle 
 
    if angle > 15: 
        return False
      
    if (height == 0 or width == 0): 
        return False
 
    area = width * height 
      
    if not self.preRatioCheck(area, width, height): 
        return False
      
    else: 
        return True
 
def preRatioCheck(self, area, width, height): 
      
    min = self.min_area 
    max = self.max_area 
 
    ratioMin = 2.5
    ratioMax = 7
 
    ratio = float(width) / float(height) 
      
    if ratio < 1: 
        ratio = 1 / ratio 
 
    if (area < min or area > max) or (ratio < ratioMin or ratio > ratioMax): 
        return False
      
    return True

Now find the contours in the validated region and validate the side ratios and area of the bounding rectangle of the largest contour in that region. After validating you will get a perfect contour of a license plate. Now extract that contour from the original image. You will get the image of plate:

Code: This step is performed by clean_plate and ratioCheck method of class PlateFinder.

def clean_plate(self, plate): 
  
    gray = cv2.cvtColor(plate, cv2.COLOR_BGR2GRAY) 
    thresh = cv2.adaptiveThreshold(gray, 255,  
                                   cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
                                   cv2.THRESH_BINARY, 11, 2) 
      
    _, contours, _ = cv2.findContours(thresh.copy(),  
                                      cv2.RETR_EXTERNAL, 
                                      cv2.CHAIN_APPROX_NONE) 
 
    if contours: 
          
        areas = [cv2.contourArea(c) for c in contours] 
          
        # index of the largest contour in the  
        # areas array 
        max_index = np.argmax(areas)   
 
        max_cnt = contours[max_index] 
        max_cntArea = areas[max_index] 
        x, y, w, h = cv2.boundingRect(max_cnt) 
 
        if not self.ratioCheck(max_cntArea, 
                               plate.shape[1], 
                               plate.shape[0]): 
            return plate, False, None
          
        return plate, True, [x, y, w, h] 
      
    else: 
        return plate, False, None
 
def ratioCheck(self, area, width, height): 
      
    min = self.min_area 
    max = self.max_area 
 
    ratioMin = 3
    ratioMax = 6
 
    ratio = float(width) / float(height) 
      
    if ratio < 1: 
        ratio = 1 / ratio 
 
    if (area < min or area > max) or (ratio < ratioMin or ratio > ratioMax): 
        return False
      
    return True

To recognize the characters on license plate precisely, we have to apply image segmentation. For that first step is to extract the value channel from the HSV format of the plate’s image. It would look like-

Now apply adaptive thresholding on the plate’s value channel image to binarize it and reveal the characters. The image of plate can have different lightning conditions in different areas, in that case adaptive thresholding can be more suitable to binarize because it uses different threshold values for different regions based on the brightness of the pixels in the region around it.

After binarizing apply bitwise not operation on the image to find the connected components in the image so that we can extract character candidates.

Construct a mask to display all the character components and then find contours in mask. After extracting the contours take the largest one, find its bounding rectangle and validate side ratios.
After validating the side ratios find the convex hull of the contour ad draw it on the character candidate mask. The mask would look like-

Now find all the contours in the character candidate mask and extract those contour areas from the plate’s value thresholded image, you will get all the characters separately.

Steps 8 to 13 are performed by segment_chars function that you can find below in the full source code. The driver code for the functions used in steps 6 to 13 is written in the method check_plate of class PlateFinder.
Now use OCR to recognize the character one by one.

Full Source Code with its working: First, create a PlateFinder class that finds the license plates and validates its size ratio and area.

import cv2 
import numpy as np 
from skimage.filters import threshold_local 
import tensorflow as tf 
from skimage import measure 
import imutils 
  
  
def sort_cont(character_contours): 
    """ 
    To sort contours 
    """
    i = 0
    boundingBoxes = [cv2.boundingRect(c) for c in character_contours] 
      
    (character_contours, boundingBoxes) = zip(*sorted(zip(character_contours, 
                                                          boundingBoxes), 
                                                      key = lambda b: b[1][i], 
                                                      reverse = False)) 
      
    return character_contours 
  
  
def segment_chars(plate_img, fixed_width): 
      
    """ 
    extract Value channel from the HSV format 
    of image and apply adaptive thresholding 
    to reveal the characters on the license plate 
    """
    V = cv2.split(cv2.cvtColor(plate_img, cv2.COLOR_BGR2HSV))[2] 
  
    thresh = cv2.adaptiveThreshold(value, 255,  
                                   cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
                                   cv2.THRESH_BINARY,  
                                   11, 2) 
  
    thresh = cv2.bitwise_not(thresh) 
  
    # resize the license plate region to 
    # a canoncial size 
    plate_img = imutils.resize(plate_img, width = fixed_width) 
    thresh = imutils.resize(thresh, width = fixed_width) 
    bgr_thresh = cv2.cvtColor(thresh, cv2.COLOR_GRAY2BGR) 
  
    # perform a connected components analysis  
    # and initialize the mask to store the locations 
    # of the character candidates 
    labels = measure.label(thresh, neighbors = 8, background = 0) 
  
    charCandidates = np.zeros(thresh.shape, dtype ='uint8') 
  
    # loop over the unique components 
    characters = [] 
    for label in np.unique(labels): 
          
        # if this is the background label, ignore it 
        if label == 0: 
            continue
        # otherwise, construct the label mask to display 
        # only connected components for the current label, 
        # then find contours in the label mask 
        labelMask = np.zeros(thresh.shape, dtype ='uint8') 
        labelMask[labels == label] = 255
  
        cnts = cv2.findContours(labelMask,  
                     cv2.RETR_EXTERNAL,  
                     cv2.CHAIN_APPROX_SIMPLE) 
  
        cnts = cnts[0] if imutils.is_cv2() else cnts[1] 
  
        # ensure at least one contour was found in the mask 
        if len(cnts) > 0: 
  
            # grab the largest contour which corresponds  
            # to the component in the mask, then grab the 
            # bounding box for the contour 
            c = max(cnts, key = cv2.contourArea) 
            (boxX, boxY, boxW, boxH) = cv2.boundingRect(c) 
  
            # compute the aspect ratio, solodity, and  
            # height ration for the component 
            aspectRatio = boxW / float(boxH) 
            solidity = cv2.contourArea(c) / float(boxW * boxH) 
            heightRatio = boxH / float(plate_img.shape[0]) 
  
            # determine if the aspect ratio, solidity,  
            # and height of the contour pass the rules 
            # tests 
            keepAspectRatio = aspectRatio < 1.0
            keepSolidity = solidity > 0.15
            keepHeight = heightRatio > 0.5 and heightRatio < 0.95
  
            # check to see if the component passes 
            # all the tests 
            if keepAspectRatio and keepSolidity and keepHeight and boxW > 14: 
                  
                # compute the convex hull of the contour 
                # and draw it on the character candidates 
                # mask 
                hull = cv2.convexHull(c) 
  
                cv2.drawContours(charCandidates, [hull], -1, 255, -1) 
  
    _, contours, hier = cv2.findContours(charCandidates, 
                                         cv2.RETR_EXTERNAL, 
                                         cv2.CHAIN_APPROX_SIMPLE) 
      
    if contours: 
        contours = sort_cont(contours) 
          
        # value to be added to each dimension  
        # of the character 
        addPixel = 4  
        for c in contours: 
            (x, y, w, h) = cv2.boundingRect(c) 
            if y > addPixel: 
                y = y - addPixel 
            else: 
                y = 0
            if x > addPixel: 
                x = x - addPixel 
            else: 
                x = 0
            temp = bgr_thresh[y:y + h + (addPixel * 2), 
                              x:x + w + (addPixel * 2)] 
  
            characters.append(temp) 
              
        return characters 
      
    else: 
        return None
  
  
  
class PlateFinder: 
    def __init__(self): 
          
        # minimum area of the plate 
        self.min_area = 4500  
          
        # maximum area of the plate 
        self.max_area = 30000  
  
        self.element_structure = cv2.getStructuringElement( 
                              shape = cv2.MORPH_RECT, ksize =(22, 3)) 
  
    def preprocess(self, input_img): 
          
        imgBlurred = cv2.GaussianBlur(input_img, (7, 7), 0) 
          
        # convert to gray 
        gray = cv2.cvtColor(imgBlurred, cv2.COLOR_BGR2GRAY)  
          
        # sobelX to get the vertical edges 
        sobelx = cv2.Sobel(gray, cv2.CV_8U, 1, 0, ksize = 3)   
          
        # otsu's thresholding 
        ret2, threshold_img = cv2.threshold(sobelx, 0, 255, 
                         cv2.THRESH_BINARY + cv2.THRESH_OTSU)  
  
        element = self.element_structure 
        morph_n_thresholded_img = threshold_img.copy() 
        cv2.morphologyEx(src = threshold_img,  
                         op = cv2.MORPH_CLOSE, 
                         kernel = element, 
                         dst = morph_n_thresholded_img) 
          
        return morph_n_thresholded_img 
  
    def extract_contours(self, after_preprocess): 
          
        _, contours, _ = cv2.findContours(after_preprocess,  
                                          mode = cv2.RETR_EXTERNAL, 
                                          method = cv2.CHAIN_APPROX_NONE) 
        return contours 
  
    def clean_plate(self, plate): 
          
        gray = cv2.cvtColor(plate, cv2.COLOR_BGR2GRAY) 
        thresh = cv2.adaptiveThreshold(gray, 
                                       255,  
                                       cv2.ADAPTIVE_THRESH_GAUSSIAN_C,  
                                       cv2.THRESH_BINARY, 
                                       11, 2) 
          
        _, contours, _ = cv2.findContours(thresh.copy(),  
                                          cv2.RETR_EXTERNAL, 
                                          cv2.CHAIN_APPROX_NONE) 
  
        if contours: 
            areas = [cv2.contourArea(c) for c in contours] 
              
            # index of the largest contour in the area 
            # array 
            max_index = np.argmax(areas)   
  
            max_cnt = contours[max_index] 
            max_cntArea = areas[max_index] 
            x, y, w, h = cv2.boundingRect(max_cnt) 
            rect = cv2.minAreaRect(max_cnt) 
              
            if not self.ratioCheck(max_cntArea, plate.shape[1],  
                                                plate.shape[0]): 
                return plate, False, None
              
            return plate, True, [x, y, w, h] 
          
        else: 
            return plate, False, None
  
  
  
    def check_plate(self, input_img, contour): 
          
        min_rect = cv2.minAreaRect(contour) 
          
        if self.validateRatio(min_rect): 
            x, y, w, h = cv2.boundingRect(contour) 
            after_validation_img = input_img[y:y + h, x:x + w] 
            after_clean_plate_img, plateFound, coordinates = self.clean_plate( 
                                                        after_validation_img) 
              
            if plateFound: 
                characters_on_plate = self.find_characters_on_plate( 
                                              after_clean_plate_img) 
                  
                if (characters_on_plate is not None and len(characters_on_plate) == 8): 
                    x1, y1, w1, h1 = coordinates 
                    coordinates = x1 + x, y1 + y 
                    after_check_plate_img = after_clean_plate_img 
                      
                    return after_check_plate_img, characters_on_plate, coordinates 
          
        return None, None, None
  
  
  
    def find_possible_plates(self, input_img): 
          
        """ 
        Finding all possible contours that can be plates 
        """
        plates = [] 
        self.char_on_plate = [] 
        self.corresponding_area = [] 
  
        self.after_preprocess = self.preprocess(input_img) 
        possible_plate_contours = self.extract_contours(self.after_preprocess) 
  
        for cnts in possible_plate_contours: 
            plate, characters_on_plate, coordinates = self.check_plate(input_img, cnts) 
              
            if plate is not None: 
                plates.append(plate) 
                self.char_on_plate.append(characters_on_plate) 
                self.corresponding_area.append(coordinates) 
  
        if (len(plates) > 0): 
            return plates 
          
        else: 
            return None
  
    def find_characters_on_plate(self, plate): 
  
        charactersFound = segment_chars(plate, 400) 
        if charactersFound: 
            return charactersFound 
  
    # PLATE FEATURES 
    def ratioCheck(self, area, width, height): 
          
        min = self.min_area 
        max = self.max_area 
  
        ratioMin = 3
        ratioMax = 6
  
        ratio = float(width) / float(height) 
          
        if ratio < 1: 
            ratio = 1 / ratio 
          
        if (area < min or area > max) or (ratio < ratioMin or ratio > ratioMax): 
            return False
          
        return True
  
    def preRatioCheck(self, area, width, height): 
          
        min = self.min_area 
        max = self.max_area 
  
        ratioMin = 2.5
        ratioMax = 7
  
        ratio = float(width) / float(height) 
          
        if ratio < 1: 
            ratio = 1 / ratio 
  
        if (area < min or area > max) or (ratio < ratioMin or ratio > ratioMax): 
            return False
          
        return True
  
    def validateRatio(self, rect): 
        (x, y), (width, height), rect_angle = rect 
  
        if (width > height): 
            angle = -rect_angle 
        else: 
            angle = 90 + rect_angle 
  
        if angle > 15: 
            return False
          
        if (height == 0 or width == 0): 
            return False
  
        area = width * height 
          
        if not self.preRatioCheck(area, width, height): 
            return False
        else: 
            return True

Here is the explanation of each and every method of PlateFinder class.
In preprocess method, following step has been done:

Blur the Image
Convert to Grayscale
Find vertical edges
Threshold the vertical edged image.
Close Morph the Threshold image.

Method extract_contours returns all external contours from the preprocessed image.
Method find_possible_plates precprocess the image with preprocess method then extracts contours by extract_contours method then it checks side ratios and area of all extracted contours and cleans the image inside the contour with check_plate and clean_plate methods. After cleaning the contour image with clean_plate method, it finds all characters on the plate with find_characters_on_plate method.
find_characters_on_plate method uses segment_chars function to find the characters. It finds characters by computing the convex hull of the contours of a thresholded value image and drawing it on the characters to reveal them.
Code: Make another class to initialize Neural Network to predict the characters on the extracted license plate.

class OCR: 
      
    def __init__(self): 
          
        self.model_file = "./model / binary_128_0.50_ver3.pb"
        self.label_file = "./model / binary_128_0.50_labels_ver2.txt"
        self.label = self.load_label(self.label_file) 
        self.graph = self.load_graph(self.model_file) 
        self.sess = tf.Session(graph = self.graph) 
  
    def load_graph(self, modelFile): 
          
        graph = tf.Graph() 
        graph_def = tf.GraphDef() 
          
        with open(modelFile, "rb") as f: 
            graph_def.ParseFromString(f.read()) 
          
        with graph.as_default(): 
            tf.import_graph_def(graph_def) 
          
        return graph 
  
    def load_label(self, labelFile): 
        label = [] 
        proto_as_ascii_lines = tf.gfile.GFile(labelFile).readlines() 
          
        for l in proto_as_ascii_lines: 
            label.append(l.rstrip()) 
          
        return label 
  
    def convert_tensor(self, image, imageSizeOuput): 
        """ 
        takes an image and tranform it in tensor 
        """
        image = cv2.resize(image,  
                           dsize =(imageSizeOuput, 
                                  imageSizeOuput), 
                           interpolation = cv2.INTER_CUBIC) 
          
        np_image_data = np.asarray(image) 
        np_image_data = cv2.normalize(np_image_data.astype('float'), 
                                      None, -0.5, .5, 
                                      cv2.NORM_MINMAX) 
          
        np_final = np.expand_dims(np_image_data, axis = 0) 
          
        return np_final 
  
    def label_image(self, tensor): 
  
        input_name = "import / input"
        output_name = "import / final_result"
  
        input_operation = self.graph.get_operation_by_name(input_name) 
        output_operation = self.graph.get_operation_by_name(output_name) 
  
        results = self.sess.run(output_operation.outputs[0], 
                                {input_operation.outputs[0]: tensor}) 
        results = np.squeeze(results) 
        labels = self.label 
        top = results.argsort()[-1:][::-1] 
          
        return labels[top[0]] 
  
    def label_image_list(self, listImages, imageSizeOuput): 
        plate = "" 
          
        for img in listImages: 
              
            if cv2.waitKey(25) & 0xFF == ord('q'): 
                break
            plate = plate + self.label_image(self.convert_tensor(img, imageSizeOuput)) 
          
        return plate, len(plate) 

It loads the pretrained OCR model and its label file in load_graph and load_label functions. label_image_list method transforms the image to tensor with convert_tensor method and then predicts the label of tensor with label_image_list function and returns the license number.
Code: Create a main function to perform the whole task in a sequence.

if __name__ == "__main__": 
      
    findPlate = PlateFinder() 
    model = OCR() 
  
    cap = cv2.VideoCapture('test_videos / video.MOV') 
      
    while (cap.isOpened()): 
        ret, img = cap.read() 
          
        if ret == True: 
            cv2.imshow('original video', img) 
              
            if cv2.waitKey(25) & 0xFF == ord('q'): 
                break
              
            possible_plates = findPlate.find_possible_plates(img) 
              
            if possible_plates is not None: 
                  
                for i, p in enumerate(possible_plates): 
                    chars_on_plate = findPlate.char_on_plate[i] 
                    recognized_plate, _ = model.label_image_list( 
                               chars_on_plate, imageSizeOuput = 128) 
  
                    print(recognized_plate) 
                    cv2.imshow('plate', p) 
                      
                    if cv2.waitKey(25) & 0xFF == ord('q'): 
                        break
        else: 
            break
              
    cap.release() 
    cv2.destroyAllWindows() 

You can download the source code with OCR model and testing video from my GitHub .
How to improve the model?

You can set a particular small region in the frame to find the plates inside it.(make sure all vehicle must pass through that region).
You can train your own machine learning model to recognize characters because the given model doesn’t recognize all the alphabets.

References:
Automatic Number Plate Recognition System (ANPR): A Survey by Chirag Indravadanbhai Patel.
Image preprocessing techniques in OpenCV documentation.

Attention reader! Don’t stop learning now. Get hold of all the important CS Theory concepts for SDE interviews with the CS Theory Course at a student-friendly price and become industry ready.

My Personal Notes

Recommended Posts: