Bld_FeatureCrps.py

#-------------------------------------------------------------------------------
# Name:        Data Preparation
# Purpose:
#
# Author:      Sardhendu_Mishra
#
# Created:     08/03/2015
# Copyright:   (c) Sardhendu_Mishra 2015
# Licence:     <your licence>
#-------------------------------------------------------------------------------

'''
    About: This code snippet prepares the features. It goes to the file where all the pictures of licence plates and non licence plate are stashed, it then does the following.
    1. Gets each picture one by one
    2. Binarizes the images, resize them, perform morphological operation. 
    3. Then calls the HOG method, creates the histogram of gradient for the each image which in our case is the feature to our classifier.
       This code snippet also prepares the class labels (1 or 0). 1 for the licence plate images and 0 for the non licence plate images.
'''

import cv2
import glob
import numpy as np
import mahotas.thresholding
import Tools
from skimage.feature import hog

import Configuration

'''
    About: The below code is the HOG class that instantiates HOG. HOG is used to create the features that is sent into a logistic regression machine or SVM Machine.
    With the use of HOG feature we develop a model or find the optimal theta value which is again used for a new instance to predict the output if the image is a number plate or not number plate
'''




class HOG:
    def __init__(self, orientations = 9, pixelsPerCell = (8, 8),cellsPerBlock = (3, 3), visualise=False, normalize = False): 
        self.orienations = orientations
        self.pixelsPerCell = pixelsPerCell
        self.cellsPerBlock = cellsPerBlock
        self.visualise = visualise
        self.normalize = normalize

    def describe(self, image):
        hist , hog_image= hog(image,
                            orientations = self.orienations,
                            pixels_per_cell = self.pixelsPerCell,
                            cells_per_block = self.cellsPerBlock,
                            visualise= self.visualise)
                            # normalise = self.normalize)  

        return hist, hog_image



#==============================================================================
#  Create data set
#==============================================================================

class CrtFeatures():    

    def __init__(self):
        self.conf = Configuration.get_datamodel_storage_path()
        self.trn_path = [self.conf['Train_data1_dir'],self.conf['Train_data2_dir']]

    def hog_feature(self, image_orig):
        # print ('CrtFeature!! hog_feature')
        #hog= a_HOG_computation.HOG(orientations=18, pixelsPerCell=(9,9), cellsPerBlock=(3,3), visualise=True, normalize=True) 
        #hog= a_HOG_computation.HOG(orientations=9, pixelsPerCell=(6,6), cellsPerBlock=(1,1), visualise=True, normalize=True)     
        hog = HOG(orientations=9, pixelsPerCell=(9,9), cellsPerBlock=(3,3), visualise=True, normalize=True)
        # Convert the coloured image into grayscale image
        image_resized = Tools.resize(image_orig, width=300)
        image_gray = cv2.cvtColor(image_resized, cv2.COLOR_BGR2GRAY)
        # Do thresholding
        image_thresh = image_gray
        T = mahotas.thresholding.otsu(image_gray) # will find an optimal value of T from the image

        image_thresh[image_thresh>T] = 255 # This goes pixel by pixel if the pixel value of the thresh is greater than the optimal value then the color is white
        image_thresh[image_thresh<T] = 0   # This goes pixel by pixel if the pixel value of the thresh is greater than the optimal value then the color is Black
        image_thresh = cv2.bitwise_not(image_thresh)

        # Perform erotion (Morphological operation)
        kernel = np.ones((2,2),np.uint8)
        image_eroded = cv2.erode(image_thresh,kernel,iterations = 1)

        # We resize the numberplate Since we are using only cars we would resize it to 60*120 pixels
        image_resized = Tools.resize(image_eroded,width=90, height=30)
        # We use HOG for licence plate detection
        hist, hog_image = hog.describe(image_resized)  
        return hist,hog_image


    def create_features(self, path):    # creates feature-set for all the images in the directory, One image is a row of the feature_matrix
        feature_arr = []
        label_arr = []
        for folder in path:
            #print folder
            for file in glob.glob(folder):
                image_orig = cv2.imread(file)
                hist,hog_image = self.hog_feature(image_orig)
                feature_arr.append(hist)
                # We prepare the class for the photos that are actual licence plate as 1 and non licence plate as 0
                if folder == path[0]:  # path 0
                    label_arr.append([1])
                else:
                    label_arr.append([0])
        return feature_arr, label_arr


    def create_training_feature(self, store=None):
        feature_trn, labels_trn = self.create_features(self.trn_path)     # Prepare the training dataset
        self.features = np.array(feature_trn, dtype="float64")
        self.labels = np.array(labels_trn)
        print (self.features.shape, self.labels.shape)

        if store:
            self.store_feature_matrix(self.conf['Data_feature_dir'], self.conf['Class_labels_dir'])

    def store_feature_matrix(self, feature_dir, label_dir):
        # We shall load the feature_train set in the disk to let our crossvalidation and classification data use the array and predict there class
        np.savetxt(feature_dir, self.features, delimiter=",")
        np.savetxt(label_dir, self.labels, delimiter=",")