Improving the robustness of all three algorithms.

Samo Penic

2018-11-17 d5c694ac711ca3b434bf16bd920b90d1a7e758c4

 sid_process.py

@@ -1,8 +1,6 @@
import cv2
import numpy as np
from skimage import morphology, img_as_ubyte
from sklearn import svm
from sklearn.externals import joblib


"""
@@ -59,6 +57,13 @@
def kernel(x, y):
    return np.ones((x, y), np.uint8)

def sid_compare(sid_no, sid_mask):
    for s,es in zip(sid_mask,sid_no):
        if s!='x' and s!=es:
            return False
    return True



def segment_by_contours(image, sorted_ctrs, classifier):
    sid_no = ""
@@ -79,23 +84,34 @@
    return sid_no


def segment_by_sid_len(image,sid_len, classifier):
    sid_no=""
    #find biggest block of pixels

    image1=cv2.morphologyEx(image,cv2.MORPH_DILATE, kernel(5,25), iterations=3)
    cv2.imwrite("sidblock1.png",image1)
def segment_by_sid_len(image, original_image, sid_mask, classifier):
    sid_no = ""
    sid_len = len(sid_mask)
    if sid_mask[0] == "1":
        move_left = 45
    elif sid_mask[0] == "x":
        move_left = 55
    else:
        move_left = 0
        # Remove noise
    image2 = cv2.morphologyEx(original_image, cv2.MORPH_OPEN, kernel(2, 2), iterations=7)
    # find biggest block of pixels
    image1 = cv2.morphologyEx(image2, cv2.MORPH_DILATE, kernel(5, 25), iterations=4)
    image1=img_as_ubyte(image1>50)
    cv2.imwrite("sidblock1.png", image1)
    im2, ctrs, hier = cv2.findContours(
        image1.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
    )
    sorted_ctrs = sorted(ctrs, key=lambda ctr: cv2.contourArea(ctr)) #get bigges contour
    sorted_ctrs = sorted(
        ctrs, key=lambda ctr: cv2.contourArea(ctr)
    )  # get bigges contour
    x, y, w, h = cv2.boundingRect(sorted_ctrs[-1])
    image=image[y:y+h,x+25:x+w-25]
    cv2.imwrite("sidblock2.png",image)
    image = image[y : y + h, x + 25 - move_left : x + w - 40] #+25,-25
    cv2.imwrite("sidblock2.png", image)
    imgHeight, imgWidth = image.shape[0:2]
    numWidth=int(imgWidth/(sid_len))
    for i in range(0,sid_len):
        num=image[:,i*numWidth:(i+1)*numWidth]
    numWidth = int(imgWidth / (sid_len))
    for i in range(0, sid_len):
        num = image[:, i * numWidth : (i + 1) * numWidth]
        num = img_as_ubyte(num < 128)
        num = cv2.resize(num, (32, 32))

@@ -104,18 +120,66 @@
        sid_no = sid_no + str(classifier.predict(num.reshape(1, -1) / 255.0)[0])
    return sid_no

def segment_by_7segments(image,original_image,sid_mask,classifier):
    block_image = cv2.morphologyEx(original_image, cv2.MORPH_CLOSE, kernel(2, 2), iterations=10)
    block_image =img_as_ubyte(block_image<50)
    cv2.imwrite("sid_3rd1.png", block_image)
    template = cv2.imread("template-8.png", 0)
    w, h = template.shape[::-1]
    res = cv2.matchTemplate(block_image, template, cv2.TM_CCOEFF_NORMED)
    loc = np.where(res >= 0.75)
    cimg = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
    loc_filtered_x=[]
    loc_filtered_y=[]
    for pt in zip(*loc[::-1]):
        pt=(pt[0]-10,pt[1]-10)
        loc_filtered_y.append(pt[1])
        loc_filtered_x.append(pt[0])
#        points.append(pt)
    #filter points
    if(len(loc_filtered_x)==0):
        return ""
    loc_filtered_x, loc_filtered_y = zip(
                *sorted(zip(loc_filtered_x, loc_filtered_y))
            )
    a = np.diff(loc_filtered_x) > int(w/2)
    a = np.append(a, True)
    loc_filtered_x = np.array(loc_filtered_x)
    loc_filtered_y = np.array(loc_filtered_y)
    points = [loc_filtered_y[a], loc_filtered_x[a]]
    for pt in zip(*points[::-1]):
        cv2.rectangle(cimg, pt, (pt[0] + w, pt[1] + h), (0, 255, 255), 2)
    cv2.imwrite("sid_3rd2.png", cimg)

    sid_no=""
    for i,pt in enumerate(zip(*points[::-1])):
        num=image[pt[1]:pt[1] + h, pt[0]:pt[0]+w]
        #cv2.imwrite("sid_3no_{}.png".format(i), num)
        num = img_as_ubyte(num < 128)
        try:
            num = cv2.resize(num, (32, 32))
        except:
            return ""
        cv2.imwrite("sid_3no_{}.png".format(i), num)
        sid_no = sid_no + str(classifier.predict(num.reshape(1, -1) / 255.0)[0])

    return sid_no

def getSID(image, classifier, sid_mask):
    sid_warn = []
    sid_err=[]
    image = 255 - image
    image_original=image.copy()
    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)
    image = cv2.morphologyEx(image, cv2.MORPH_OPEN, kernel(2, 2), iterations=3)
    # Closing. Connect non connected parts
    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)
    # image = cv2.morphologyEx(image, cv2.MORPH_OPEN, kernel(8, 8), iterations=1)
    # don't do too much noise removal.
    image = cv2.morphologyEx(image, cv2.MORPH_OPEN, kernel(3, 3), iterations=1)

    # Skeletonization
@@ -129,21 +193,27 @@
    # Thining again
    image = img_as_ubyte(morphology.skeletonize(image > 0.5))
    image = cv2.morphologyEx(image, cv2.MORPH_DILATE, kernel(10, 10))
    cv2.imwrite("enhancedSID.png",image)
    cv2.imwrite("enhancedSID.png", image)
    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])

    sid_no = ""
    #sid_len = len(sid_mask)
    #sid_no = segment_by_sid_len(image, sid_len, classifier)
    #if sid_mask is not None:
    print(len(sid_mask),len(sorted_ctrs))
    #if len(sid_mask)==len(sorted_ctrs):
    sid_no=segment_by_contours(image,sorted_ctrs[1:],classifier)
    print(len(sid_mask), len(sorted_ctrs))
    sid_no = segment_by_contours(
        image, sorted_ctrs[1:], classifier
    )  # we remove largest contour that surrounds whole image
    print(sid_no)
    if(len(sid_no)!=len(sid_mask)):
        print("Ooops have to find another way")
        sid_no=segment_by_sid_len(image,len(sid_mask),classifier)
    return sid_no
    if len(sid_no) != len(sid_mask) or not sid_compare(sid_no,sid_mask):
        sid_warn.append("Trying second SID algorithm.")
        sid_no = segment_by_7segments(image, image_original, sid_mask, classifier)
    print(sid_no)
    if(len(sid_no))!=len(sid_mask):
        sid_no = segment_by_sid_len(image, image_original, sid_mask, classifier)
        sid_warn.append("Trying third SID algorithm.")


    if not sid_compare(sid_no, sid_mask):
            sid_err=['Wrong SID!']

    return (sid_no, sid_err, sid_warn)