def crack(tocrack):
im = cv.CreateImage(cv.GetSize(tocrack), 8, 1)
cv.Split(tocrack, None, None, im, None)
cv.Threshold(im, im, 250, 255, cv.CV_THRESH_BINARY)
txt = pytesser.iplimage_to_string(im)
return txt[:-2]
def crack(tocrack):
im = cv.CreateImage(cv.GetSize(tocrack), 8, 1)
cv.Split(tocrack, None, None, im, None)
cv.Threshold(im, im, 250, 255, cv.CV_THRESH_BINARY)
txt = pytesser.iplimage_to_string(im)
return txt[:-2]
def process_all(results):
dir = "Ebay" #Consider that all images are stored in the dir 'Ebay'
for file,r in zip(os.listdir(dir),results):
im = cv.LoadImage(os.path.join(dir,file),cv.CV_LOAD_IMAGE_GRAYSCALE) #Load the image
im = crack(im) #intent to crack it
res = pytesser.iplimage_to_string(im,psm=pytesser.PSM_SINGLE_WORD) #Do characters recognition
res = res[:-2] #Remove the two \n\n always put at the end of the result
if res == r: #Compare the result of the value contained in our list
print file+": "+res+" | "+r+ " OK"
else:
print file+": "+res+" | "+r+" NO"
def crack(self,dilateiter=4, erodeiter=4, threshold=200, size=(155,55)): #Take all parameters
resized = resizeImage(self.image, (self.image.width*6, self.image.height*6))
dilateImage(resized, dilateiter)
erodeImage(resized, erodeiter)
thresholdImage(resized,threshold, cv.CV_THRESH_BINARY)
resized = resizeImage(resized, size)
#Call the tesseract engine
ret = pytesser.iplimage_to_string(resized)
ret = ret[:-2]
return ret
def crack(self,
dilateiter=4,
erodeiter=4,
threshold=200,
size=(155, 55)): #Take all parameters
resized = resizeImage(self.image,
(self.image.width * 6, self.image.height * 6))
dilateImage(resized, dilateiter)
erodeImage(resized, erodeiter)
thresholdImage(resized, threshold, cv.CV_THRESH_BINARY)
resized = resizeImage(resized, size)
#Call the tesseract engine
ret = pytesser.iplimage_to_string(resized)
ret = ret[:-2]
return ret
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret, th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
h, w = img.shape
img = cv2.resize(img, (w * resize, h * resize),
interpolation=cv2.INTER_CUBIC)
return img
def feature(img):
kernel = np.ones((1, 1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open,
cv2.MORPH_CLOSE,
kernel,
iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img):
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
ft = feature(resize(normalize(img)))
text = pytesser.iplimage_to_string(cv.fromarray(ft), 'eng').strip()
if self._debug:
debug(ft)
text = raw_input("debug:")
return text if text else ''
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret,th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
h,w = img.shape
img = cv2.resize(img, (w*resize,h*resize), interpolation=cv2.INTER_CUBIC)
return img
def feature(img):
kernel = np.ones((1,1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open, cv2.MORPH_CLOSE, kernel, iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img):
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
ft = feature(resize(normalize(img)))
text = pytesser.iplimage_to_string(cv.fromarray(ft), 'eng').strip()
if self._debug:
debug(ft)
text = raw_input("debug:")
return text if text else ''
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret,th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
h,w = img.shape
img = cv2.resize(img, (w*resize,h*resize), interpolation=cv2.INTER_CUBIC)
return img
def boundary(img, bund=3):
h, w = img.shape
# find best match captcha area
# smooth background noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
contours,hierarchy = cv2.findContours(th1, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
area = lambda (x, y, w, h): (w*h, x, y, w, h)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours], reverse=True)
# iter sub captcha
x0 = lambda x: x-bund if x > bund else 0
y0 = lambda y: y-bund if y > bund else 0
x1 = lambda x: x+bund if w-bund > x else w
y1 = lambda y: y+bund if h-bund > y else h
for it in sorted(best[:5], key=lambda x: x[1]):
yield ((x0(it[1]),y0(it[2])), (x1(it[1]+it[3]),y1(it[2]+it[4])))
def feature(img):
kernel = np.ones((1,1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open, cv2.MORPH_CLOSE, kernel, iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img, bund):
cv2.rectangle(img, bund[0], bund[1], (255,255,255), 1)
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
text = ''
img = resize(normalize(img), 1)
h, w = img.shape
bkimg = np.zeros((h,w,3), np.uint8)
bkimg = resize(normalize(bkimg), 1)
for bund in boundary(img):
x,y = zip(*bund)
bkimg[y[0]:y[1],x[0]:x[1]] = img[y[0]:y[1],x[0]:x[1]]
if self._debug:
debug(feature(bkimg), bund)
text = pytesser.iplimage_to_string(cv.fromarray(feature(bkimg)), 'eng').strip()
return text if text else ''
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
# rgb 2 gray
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret,th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
# scalar must be int
h,w = img.shape
img = cv2.resize(img, (w*resize,h*resize), interpolation=cv2.INTER_CUBIC)
return img
def img_absdiff(imgA, imgB):
xA, yA = imgA.shape
xB, yB = imgB.shape
# resize imgB to imgA if size not same
if xB != xA or yB != yA:
imgB = cv2.resize(imgB, (x,y), interpolation=cv2.INTER_CUBIC)
diff = np.diff(imgA.astype('float') - imgB.astype('float'))
absdiff = np.sum(diff**2)
return absdiff / (xA * yA)
def feature_extract(img):
# smooth background noise
kernel = np.ones((4,4), np.uint8)
erosion = cv2.erode(img, kernel, iterations=1)
blurred = cv2.GaussianBlur(erosion, (5,5), 0)
# enhance img edge
edged = cv2.Canny(blurred, 30, 150)
dilation = cv2.dilate(edged, kernel, iterations=1)
return dilation
def ini_img_as_blank(h, w):
bkimg = np.zeros((h,w,3), np.uint8)
bkimg = resize(normalize(bkimg), 1)
return bkimg
def bitwise_and_noise_reduce(img):
kernel = np.ones((1,1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open, cv2.MORPH_CLOSE, kernel, iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def iter_find_best_char_index(img, bund=1, limit=5):
# feature select order by roll win area size
h, w = img.shape
area = lambda (x, y, w, h): (w*h, x, y, w, h)
contours, hierarchy = cv2.findContours(img.copy(), cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours], reverse=True)
# iter sub captcha
x0 = lambda x: x-bund if x > bund else 0
y0 = lambda y: y-bund if y > bund else 0
x1 = lambda x: x+bund if w-bund > x else w
y1 = lambda y: y+bund if h-bund > y else h
for it in sorted(best[:limit], key=lambda x: x[1]):
yield ((x0(it[1]),y0(it[2])), (x1(it[1]+it[3]),y1(it[2]+it[4])))
def blank_boundary_edge(img, outer=4):
h, w = img.shape
bh = h+outer*2
bw = w+outer*2
bkimg = ini_img_as_blank(bh,bw)
bkimg[outer:h+outer,outer:w+outer] = img[0:h,0:w]
return bkimg
def debug(img):
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
chars = []
nwimg = bitwise_and_noise_reduce(normalize(img))
h, w = nwimg.shape
for bund in iter_find_best_char_index(nwimg, 1, 5):
x,y = zip(*bund)
ff = nwimg[y[0]:y[1],x[0]:x[1]]
nwff = blank_boundary_edge(ff.copy(), 4)
if self._debug:
debug(nwff)
char = pytesser.iplimage_to_string(cv.fromarray(nwff), 'eng', 10).strip()
chars.append(char.upper())
return ''.join(chars)
def detect(filename, folder, file_no):
# print 'in'
img = cv2.imread(filename)
img = cv2.medianBlur(img, 1) #smothing image
# extracting white characters
image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
imgray = crop(image)
# imgray = image
imgray = cv2.resize(imgray, (880, 100))
(h, w) = imgray.shape
(img_h, img_w) = (h, w)
drawing = noise(imgray, imgray.shape, 1000, 25000)
cv2.imshow('win1', drawing)
cv2.waitKey()
cv2.imwrite('E:\Results\\res.jpg', drawing)
# character recorgnition
image = cv.LoadImage("E:\Results\\res.jpg", cv.CV_LOAD_IMAGE_GRAYSCALE)
lang = ['eng']
result = 0
for i in lang:
data = pytesser.iplimage_to_string(image,
i,
pytesser.PSM_SINGLE_LINE,
makebox=True)
# print data
word = ''
section = [] #all the characters and coordinates in a list
line = []
for i in data:
if i == ' ' or i == '\n':
line.append(word)
word = ''
elif (i != '\n'):
word += i
if i == '\n':
section.append(line)
line = []
coord = []
chars = [] #coordinates of the characters
count = 0
for i in section:
for j in i:
# print count
if j.isdigit() and count < 5 and count > 0:
coord.append(int(j))
count += 1
count = 0
chars.append(coord)
coord = []
# removing all symbols
dele = 0
delete = []
for i in section:
if not i[0].isdigit() and not i[0].isalpha():
delete.append(dele)
dele += 1
delete.reverse()
for i in delete:
section.pop(i)
chars.pop(i)
c = 0
point = []
for i in chars:
cv2.rectangle(drawing, (i[0], i[1]), (i[2], i[3]), (255, 255, 255), 2)
cv2.imshow('rec', drawing)
cv2.waitKey(1)
cv2.imwrite('E:\\report\\res.jpg', drawing)
counter = 0
line_coordinates = [0]
if len(chars) == 1:
for i in chars:
if i[0] > 400:
line_coordinates.append(i[0] / 2)
line_coordinates.append(i[0])
# for debugging, draws box using coordinates received from make box function
for i in chars:
counter += 1
if c == 0:
point.append(i[2])
c = 1
elif c == 1:
point.append(i[0])
line = ((point[0] + point[1]) / 2)
line_coordinates.append(line)
cv2.line(drawing, (line, img_h), (line, 0), (255, 255, 255), 2)
c = 0
point = []
if counter == len(chars):
if img_w - i[2] < 150:
line_coordinates.append(img_w)
else:
line_coordinates.append(i[2])
line_coordinates.append(img_w)
cv2.line(drawing, (i[2], img_h), (i[2], 0), (255, 255, 255), 2)
# spliting image and processing
iterator = 1
id = ''
while iterator < len(line_coordinates):
roi = imgray[0:img_h,
line_coordinates[iterator - 1]:line_coordinates[iterator]]
iterator += 1
corrected = noise(roi, roi.shape, 300, 20000)
cv2.imwrite('E:\Results\\res.jpg', corrected)
cv2.imshow('win1', corrected)
cv2.waitKey(1)
image = cv.LoadImage("E:\Results\\res.jpg", cv.CV_LOAD_IMAGE_GRAYSCALE)
data = pytesser.iplimage_to_string(image, 'enm', 7)
# print data
for i in data:
if i == '!':
i = '1'
if i.isalpha() or i.isdigit():
if i == 'O' or i == 'o':
i = '0'
if i == 'L' or i == 'l':
i = '1'
if i == 'i' or i == 'I':
i = '1'
id = id + i
# print id
if id != '':
print id
sol = check(id)
final = sol[0]
alternative = sol[1]
possible = sol[2]
else:
return None
# print sol
if not final:
os.chdir(folder)
filename = 'error' + str(file_no)
img_name = filename + '.jpg'
txt_name = filename + '.txt'
cv2.imwrite(img_name, img)
if len(alternative) > 1: #storing alternatives in a text file
f = open(txt_name, 'w')
for i in alternative:
f.write(i + '\n')
f.close()
return None
else:
return possible
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret, th0 = cv2.threshold(gray, 150, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=2):
h, w = img.shape
img = cv2.resize(img, (w * resize, h * resize),
interpolation=cv2.INTER_CUBIC)
return img
def filter(img):
edges = cv2.Canny(img, 150, 250, apertureSize=3)
lines = cv2.HoughLinesP(edges, 1, np.pi / 180, 1, 3, 1)
# remove boundary as white line
for line in lines:
for x1, y1, x2, y2 in line:
cv2.line(img, (x1, y1), (x2, y2), (255, 255, 255), 14)
# smooth background noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 250, 255, cv2.THRESH_BINARY)
return th1
def extend(img, ext=30):
h, w = img.shape
wtimg = 255 - normalize(np.zeros((h + ext, w, 3), np.uint8))
wtimg[(h + ext) // 2 - h // 2:(h + ext) // 2 +
h // 2, :w] = img[:h, :w]
return wtimg
def boundary(img, bund=-1):
# find best match captcha area
contours, hierarchy = cv2.findContours(img, cv2.RETR_TREE,
cv2.CHAIN_APPROX_SIMPLE)
area = lambda (x, y, w, h): (w * h, x, y, w, h)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours],
reverse=True)
# iter sub captcha
x0 = lambda x: x - bund if x > bund else 0
y0 = lambda y: y - bund if y > bund else 0
x1 = lambda x: x + bund if w - bund > x else w
y1 = lambda y: y + bund if h - bund > y else h
for it in sorted(best[1:6], key=lambda x: x[1]):
yield ((x0(it[1]), y0(it[2])), (x1(it[1] + it[3]),
y1(it[2] + it[4])))
def feature(img):
kernel = np.ones((1, 1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open,
cv2.MORPH_CLOSE,
kernel,
iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img, bund):
cv2.rectangle(img, bund[0], bund[1], (0, 0, 0), 1)
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
text = ''
img = extend(resize(normalize(img), 4))
h, w = img.shape
bkimg = np.zeros((h, w, 3), np.uint8)
bkimg = 255 - resize(normalize(bkimg), 1)
for bund in boundary(filter(img.copy()), -7):
x, y = zip(*bund)
bkimg[y[0]:y[1], x[0]:x[1]] = img[y[0]:y[1], x[0]:x[1]]
if self._debug:
debug(bkimg, bund)
text = pytesser.iplimage_to_string(cv.fromarray(bkimg), 'eng').strip()
return text if text else ''
upper_line=[] #stores elements from maxValueOfRowsOfEdgesImage which denotes starting of line in image
lower_line=[] #stores elements from maxValueOfRowsOfEdgesImage which denotes ending of line in image
#this loop finds at which row in image, line starts and at which row line in image ends
for i in range(rows+1):
if ((maxValueOfRowsOfEdgesImage[i+1]==0)|(maxValueOfRowsOfEdgesImage[i]==0)):
if ((maxValueOfRowsOfEdgesImage[i+1]!=0) & (maxValueOfRowsOfEdgesImage[i]==0)):
#img_deskewed = cv2.line(img_deskewed,(0,i),(cols,i),(100,0,100),1) #this line draws upper line on image
upper_line.append(i)
elif((maxValueOfRowsOfEdgesImage[i+1]==0) & (maxValueOfRowsOfEdgesImage[i]!=0)):
#img_deskewed = cv2.line(img_deskewed,(0,i),(cols,i),(100,0,100),1) #this line draws lower line on image
lower_line.append(i)
upper_line=np.array(upper_line) #converts list in array
lower_line=np.array(lower_line)
#cuts each line from image and does character recognition for each lines separately
fout = open(output+".txt","w")
print "Please wait while I'm processing your Image..."
for i in range(len(upper_line)):
h = lower_line[i]-upper_line[i] #hight of line
y = upper_line[i] #Y coordinate of line
img_line = img_deskewed[y:y+h,0:cols] #cut detected line from image and save in img_line
if(thr == True):
img_line = cv2.adaptiveThreshold(img_line,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,\
cv2.THRESH_BINARY,15,2)
txt = pt.iplimage_to_string(img_line, lang, psm) #txt contains recognized string from line
fout.write(txt)
fout.close()
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret,th0 = cv2.threshold(gray, 150, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=2):
h,w = img.shape
img = cv2.resize(img, (w*resize,h*resize), interpolation=cv2.INTER_CUBIC)
return img
def filter(img):
edges = cv2.Canny(img, 150, 250, apertureSize=3)
lines = cv2.HoughLinesP(edges,1,np.pi/180, 1, 3, 1)
# remove boundary as white line
for line in lines:
for x1,y1,x2,y2 in line:
cv2.line(img,(x1,y1),(x2,y2),(255,255,255),14)
# smooth background noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 250, 255, cv2.THRESH_BINARY)
return th1
def extend(img, ext=30):
h, w = img.shape
wtimg = 255 - normalize(np.zeros((h+ext,w,3), np.uint8))
wtimg[(h+ext)//2-h//2:(h+ext)//2+h//2,:w] = img[:h,:w]
return wtimg
def boundary(img, bund=-1):
# find best match captcha area
contours,hierarchy = cv2.findContours(img, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
area = lambda (x, y, w, h): (w*h, x, y, w, h)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours], reverse=True)
# iter sub captcha
x0 = lambda x: x-bund if x > bund else 0
y0 = lambda y: y-bund if y > bund else 0
x1 = lambda x: x+bund if w-bund > x else w
y1 = lambda y: y+bund if h-bund > y else h
for it in sorted(best[1:6], key=lambda x: x[1]):
yield ((x0(it[1]),y0(it[2])), (x1(it[1]+it[3]),y1(it[2]+it[4])))
def feature(img):
kernel = np.ones((1,1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5,5), 0)
# threshold filter
ret,th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open, cv2.MORPH_CLOSE, kernel, iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img, bund):
cv2.rectangle(img, bund[0], bund[1], (0,0,0), 1)
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
text = ''
img = extend(resize(normalize(img), 4))
h, w = img.shape
bkimg = np.zeros((h,w,3), np.uint8)
bkimg = 255 - resize(normalize(bkimg), 1)
for bund in boundary(filter(img.copy()), -7):
x,y = zip(*bund)
bkimg[y[0]:y[1],x[0]:x[1]] = img[y[0]:y[1],x[0]:x[1]]
if self._debug:
debug(bkimg, bund)
text = pytesser.iplimage_to_string(cv.fromarray(bkimg), 'eng').strip()
return text if text else ''
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret, th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
h, w = img.shape
img = cv2.resize(img, (w * resize, h * resize),
interpolation=cv2.INTER_CUBIC)
return img
def boundary(img, bund=3):
h, w = img.shape
# find best match captcha area
# smooth background noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
contours, hierarchy = cv2.findContours(th1, cv2.RETR_TREE,
cv2.CHAIN_APPROX_SIMPLE)
area = lambda (x, y, w, h): (w * h, x, y, w, h)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours],
reverse=True)
# iter sub captcha
x0 = lambda x: x - bund if x > bund else 0
y0 = lambda y: y - bund if y > bund else 0
x1 = lambda x: x + bund if w - bund > x else w
y1 = lambda y: y + bund if h - bund > y else h
for it in sorted(best[:5], key=lambda x: x[1]):
yield ((x0(it[1]), y0(it[2])), (x1(it[1] + it[3]),
y1(it[2] + it[4])))
def feature(img):
kernel = np.ones((1, 1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open,
cv2.MORPH_CLOSE,
kernel,
iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def debug(img, bund):
cv2.rectangle(img, bund[0], bund[1], (255, 255, 255), 1)
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
text = ''
img = resize(normalize(img), 1)
h, w = img.shape
bkimg = np.zeros((h, w, 3), np.uint8)
bkimg = resize(normalize(bkimg), 1)
for bund in boundary(img):
x, y = zip(*bund)
bkimg[y[0]:y[1], x[0]:x[1]] = img[y[0]:y[1], x[0]:x[1]]
if self._debug:
debug(feature(bkimg), bund)
text = pytesser.iplimage_to_string(cv.fromarray(feature(bkimg)),
'eng').strip()
return text if text else ''
def run(self, img):
def normalize(img):
img = cv2.cvtColor(img, cv2.CV_32F)
# rgb 2 gray
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
gray = cv2.equalizeHist(gray)
ret, th0 = cv2.threshold(gray, 200, 255, cv2.THRESH_BINARY)
return th0
def resize(img, resize=1):
# scalar must be int
h, w = img.shape
img = cv2.resize(img, (w * resize, h * resize),
interpolation=cv2.INTER_CUBIC)
return img
def img_absdiff(imgA, imgB):
xA, yA = imgA.shape
xB, yB = imgB.shape
# resize imgB to imgA if size not same
if xB != xA or yB != yA:
imgB = cv2.resize(imgB, (x, y), interpolation=cv2.INTER_CUBIC)
diff = np.diff(imgA.astype('float') - imgB.astype('float'))
absdiff = np.sum(diff**2)
return absdiff / (xA * yA)
def feature_extract(img):
# smooth background noise
kernel = np.ones((4, 4), np.uint8)
erosion = cv2.erode(img, kernel, iterations=1)
blurred = cv2.GaussianBlur(erosion, (5, 5), 0)
# enhance img edge
edged = cv2.Canny(blurred, 30, 150)
dilation = cv2.dilate(edged, kernel, iterations=1)
return dilation
def ini_img_as_blank(h, w):
bkimg = np.zeros((h, w, 3), np.uint8)
bkimg = resize(normalize(bkimg), 1)
return bkimg
def bitwise_and_noise_reduce(img):
kernel = np.ones((1, 1), np.uint8)
# smooth backgroun noise
blur = cv2.GaussianBlur(img, (5, 5), 0)
# threshold filter
ret, th1 = cv2.threshold(blur, 235, 255, cv2.THRESH_BINARY)
# colsing/opening
open = cv2.morphologyEx(th1, cv2.MORPH_OPEN, kernel, iterations=5)
close = cv2.morphologyEx(open,
cv2.MORPH_CLOSE,
kernel,
iterations=5)
mask = cv2.bitwise_and(th1, th1, mask=close)
return mask
def iter_find_best_char_index(img, bund=1, limit=5):
# feature select order by roll win area size
h, w = img.shape
area = lambda (x, y, w, h): (w * h, x, y, w, h)
contours, hierarchy = cv2.findContours(img.copy(), cv2.RETR_TREE,
cv2.CHAIN_APPROX_SIMPLE)
best = sorted([area(cv2.boundingRect(cnt)) for cnt in contours],
reverse=True)
# iter sub captcha
x0 = lambda x: x - bund if x > bund else 0
y0 = lambda y: y - bund if y > bund else 0
x1 = lambda x: x + bund if w - bund > x else w
y1 = lambda y: y + bund if h - bund > y else h
for it in sorted(best[:limit], key=lambda x: x[1]):
yield ((x0(it[1]), y0(it[2])), (x1(it[1] + it[3]),
y1(it[2] + it[4])))
def blank_boundary_edge(img, outer=4):
h, w = img.shape
bh = h + outer * 2
bw = w + outer * 2
bkimg = ini_img_as_blank(bh, bw)
bkimg[outer:h + outer, outer:w + outer] = img[0:h, 0:w]
return bkimg
def debug(img):
cv2.imshow('test', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
chars = []
nwimg = bitwise_and_noise_reduce(normalize(img))
h, w = nwimg.shape
for bund in iter_find_best_char_index(nwimg, 1, 5):
x, y = zip(*bund)
ff = nwimg[y[0]:y[1], x[0]:x[1]]
nwff = blank_boundary_edge(ff.copy(), 4)
if self._debug:
debug(nwff)
char = pytesser.iplimage_to_string(cv.fromarray(nwff), 'eng',
10).strip()
chars.append(char.upper())
return ''.join(chars)
def detect(filename, folder, file_no):
# print 'in'
img = cv2.imread(filename)
img = cv2.medianBlur(img, 1) #smothing image
# extracting white characters
image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
imgray = crop(image)
# imgray = image
imgray = cv2.resize(imgray, (880,100))
(h,w) = imgray.shape
(img_h,img_w) = (h,w)
drawing = noise(imgray, imgray.shape, 1000, 25000)
cv2.imshow('win1', drawing)
cv2.waitKey()
cv2.imwrite('E:\Results\\res.jpg', drawing)
# character recorgnition
image = cv.LoadImage("E:\Results\\res.jpg", cv.CV_LOAD_IMAGE_GRAYSCALE)
lang = ['eng']
result = 0
for i in lang:
data = pytesser.iplimage_to_string(image, i, pytesser.PSM_SINGLE_LINE,makebox=True)
# print data
word = ''
section = [] #all the characters and coordinates in a list
line = []
for i in data:
if i == ' ' or i=='\n':
line.append(word)
word = ''
elif (i!='\n'):
word+=i
if i == '\n':
section.append(line)
line = []
coord = []
chars = [] #coordinates of the characters
count = 0
for i in section:
for j in i:
# print count
if j.isdigit() and count < 5 and count > 0:
coord.append(int(j))
count += 1
count = 0
chars.append(coord)
coord = []
# removing all symbols
dele= 0
delete=[]
for i in section:
if not i[0].isdigit() and not i[0].isalpha():
delete.append(dele)
dele += 1
delete.reverse()
for i in delete:
section.pop(i)
chars.pop(i)
c = 0
point = []
for i in chars:
cv2.rectangle(drawing,(i[0],i[1]),(i[2],i[3]),(255,255,255),2)
cv2.imshow('rec', drawing)
cv2.waitKey(1)
cv2.imwrite('E:\\report\\res.jpg', drawing)
counter = 0
line_coordinates = [0]
if len(chars) == 1:
for i in chars:
if i[0] > 400:
line_coordinates.append(i[0]/2)
line_coordinates.append(i[0])
# for debugging, draws box using coordinates received from make box function
for i in chars:
counter += 1
if c == 0:
point.append(i[2])
c = 1
elif c == 1:
point.append(i[0])
line = ((point[0] + point[1]) / 2)
line_coordinates.append(line)
cv2.line(drawing,(line,img_h),(line,0),(255,255,255),2)
c = 0
point = []
if counter == len(chars):
if img_w - i[2] < 150:
line_coordinates.append(img_w)
else:
line_coordinates.append(i[2])
line_coordinates.append(img_w)
cv2.line(drawing,(i[2],img_h),(i[2],0),(255,255,255),2)
# spliting image and processing
iterator = 1
id = ''
while iterator < len(line_coordinates):
roi = imgray[0:img_h, line_coordinates[iterator-1]:line_coordinates[iterator]]
iterator += 1
corrected = noise(roi, roi.shape, 300, 20000)
cv2.imwrite('E:\Results\\res.jpg', corrected)
cv2.imshow('win1', corrected)
cv2.waitKey(1)
image = cv.LoadImage("E:\Results\\res.jpg", cv.CV_LOAD_IMAGE_GRAYSCALE)
data = pytesser.iplimage_to_string(image, 'enm', 7 )
# print data
for i in data:
if i == '!':
i = '1'
if i.isalpha() or i.isdigit():
if i == 'O' or i =='o':
i = '0'
if i == 'L' or i == 'l':
i = '1'
if i == 'i' or i == 'I':
i = '1'
id = id + i
# print id
if id != '':
print id
sol = check(id)
final = sol[0]
alternative = sol[1]
possible = sol[2]
else:
return None
# print sol
if not final:
os.chdir(folder)
filename = 'error' + str(file_no)
img_name = filename + '.jpg'
txt_name = filename + '.txt'
cv2.imwrite(img_name, img)
if len(alternative) > 1: #storing alternatives in a text file
f = open(txt_name, 'w')
for i in alternative:
f.write(i+'\n')
f.close()
return None
else:
return possible
