Initial version of sid recognition. Cutting the numerals is not perfect yet.

Samo Penic

2018-11-16 02e0f7bc51acfa06e7299919b54b56a3c7eca02b

Initial version of sid recognition. Cutting the numerals is not perfect yet.

 .idea/sonarIssues.xml

@@ -13,6 +13,11 @@
            <set />
          </value>
        </entry>
        <entry key="/a.dummy">
          <value>
            <set />
          </value>
        </entry>
        <entry key="$PROJECT_DIR$/Ocr.py">
          <value>
            <set />

 Ocr.py

@@ -1,17 +1,18 @@
from pyzbar.pyzbar import decode
from sid_process import enhanceSID
from sid_process import getSID
import cv2
import numpy as np
import math


class Paper:
    def __init__(self, filename=None):
    def __init__(self, filename=None, sid_classifier=None):
        self.filename = filename
        self.invalid = None
        self.QRData = None
        self.errors = []
        self.warnings = []
        self.sid_classifier=sid_classifier
        if filename is not None:
            self.loadImage(filename)
            self.runOcr()
@@ -137,8 +138,8 @@
            loc_filtered_x, loc_filtered_y = zip(
                *sorted(zip(loc_filtered_x, loc_filtered_y))
            )
        # loc=[loc_filtered_y,loc_filtered_x]
        # remove duplicates
            # loc=[loc_filtered_y,loc_filtered_x]
            # remove duplicates
            a = np.diff(loc_filtered_x) > 40
            a = np.append(a, True)
            loc_filtered_x = np.array(loc_filtered_x)
@@ -213,5 +214,13 @@
            self.answerMatrix.append(oneline)

    def get_enhanced_sid(self):
        es= enhanceSID(self.img[int(0.04*self.imgHeight):int(0.08*self.imgHeight), int(0.7*self.imgWidth):int(0.99*self.imgWidth)])
        cv2.imwrite("enhancedSID.png",es)
        if self.sid_classifier is None:
            return "x"
        es = getSID(
            self.img[
                int(0.045 * self.imgHeight) : int(0.085 * self.imgHeight),
                int(0.7 * self.imgWidth) : int(0.99 * self.imgWidth),
            ],
            self.sid_classifier,
        )
        return es

 aoiOcr.py

@@ -1,8 +1,12 @@
from Ocr import Paper
from sklearn.externals import joblib

classifier = joblib.load('filename.joblib')

#p=Paper(filename='testpage300dpi_scan1.png')
p=Paper(filename='sizif111.tif')
p=Paper(filename='sizif111.tif', sid_classifier=classifier)
#p=Paper(filename='processed_scans/20141016095134535_0028.tif')

print(p.QRData)
print(p.errors)


 sid_process.py

@@ -1,6 +1,10 @@
import cv2
import numpy as np
from skimage import morphology,img_as_ubyte
from sklearn import svm
from sklearn.externals import joblib



"""
  (1) The text is an array of chars (in row-major order) where
@@ -57,19 +61,18 @@
    return np.ones((x, y), np.uint8)


def enhanceSID(image):
def getSID(image, classifier):
    image=255-image
    image=img_as_ubyte(image>100)
    cv2.imwrite("enSID0.png", image)
    # Remove noise
    image = cv2.morphologyEx(image, cv2.MORPH_OPEN, kernel(2,2), iterations=1)
    # Closing. Connect non connected parts
    image = cv2.morphologyEx(image, cv2.MORPH_CLOSE, kernel(5, 5), iterations=2)
    image = cv2.morphologyEx(image, cv2.MORPH_CLOSE, kernel(5, 3), iterations=4)
    # Again noise removal after closing

    image = cv2.morphologyEx(image, cv2.MORPH_OPEN, kernel(8,8), iterations=1)
    # Skeletonization
    ##For thinning I am using erosion
    ##image = cv2.erode(image,kernel(4,4),iterations = 40)
    image = img_as_ubyte(morphology.thin(image>128))
    cv2.imwrite("enSID1.png",image)
    # Stub removal (might not be necessary if thinning instead of skeletonize is used above
@@ -80,4 +83,26 @@
    # Thining again
    image = img_as_ubyte(morphology.skeletonize(image>0.5))
    image = cv2.morphologyEx(image, cv2.MORPH_DILATE, kernel(10, 10))

    im2,ctrs, hier = cv2.findContours(image.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    sorted_ctrs = sorted(ctrs, key=lambda ctr: cv2.boundingRect(ctr)[0])

    #classifier = joblib.load('filename.joblib')

    sid_no=""
    for i, ctr in enumerate(sorted_ctrs):
        # Get bounding box
        x, y, w, h = cv2.boundingRect(ctr)
        # Getting ROI
        if(w<h/2):
            sid_no=sid_no+"1"
            continue
        roi = image[y:y+h, x:x+w]
        roi = img_as_ubyte(roi < 128)
        roi = cv2.resize(roi,(32,32))

        #cv2.rectangle(image,(x,y),( x + w, y + h ),(0,255,0),2)
        cv2.imwrite('sid_no_{}.png'.format(i), roi)
        sid_no=sid_no+str(classifier.predict(roi.reshape(1,-1)/255.0)[0])
    print(sid_no)
    return image