def completeClassification(image, svg_file_path):
# This function classifies the setch strokes according to the allocated stamps
draw.showImage(image)
# Detecting symbols
predictions = sd.detectSymbols(image,
symbolType='stamp',
method='t',
filterColor='red')
#print("predictions",predictions)
beacons = sd.beaconCoordinates(predictions)
#print(beacons)
#if len(beacons)>0:
beaconsSVG = getBeacons(beacons, image)
# Here we get rid of the symbols, clean and segmantate the sketch to obtain a clean image,
# Contours are then extracted from this clean image
contours = ss.segmentateSketch(image, 'skeleton') #
# To provide a better overview of the extracted contours we display the five larger ones one by one
#draw.drawContoursOneByOne(contours, image, numberOfContours=5)
# We classify the contours based on their distances with respect to the detected symbols
strokeList = strokeClassification(image, contours, predictions, knn=1)
draw.drawHandStrokes(image, strokeList, oneByOne=False)
strokesToSVG(strokeList, beaconsSVG, image, svg_file_path)
return strokeList
def segmentateSketch(image, method='skeleton'):
# Returns the main contours inside an image. In case of selecting the skeleton method
# all the sketch strokes are transformed into one pixel width lines
imageClone = image.copy()
draw.showImage(imageClone)
imageWithoutStamps = cleanSymbols(imageClone,
colorSpace='hsv',
color='red')
#draw.showImage(imageWithoutStamps)
contours = getContours(imageWithoutStamps, method)
return contours
def getSkeleton(image):
# Applies a skeletonization algorithm, thus, transforming all the strokes inside the sketch into one pixel width lines
grayscaled = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
#grayscaled = cv2.medianBlur(grayscaled, 5)
ret, threshold = cv2.threshold(grayscaled, 200, 255, cv2.THRESH_BINARY)
draw.showImage(threshold)
threshold = cv2.bitwise_not(threshold)
threshold[threshold == 255] = 1
skeleton = skeletonize(threshold)
skeleton = img_as_ubyte(skeleton)
return skeleton
import cv2
from sketchProcessor.helperLibraries.utils import draw
from sketchProcessor.helperLibraries.utils import simple_obj_det as sd
from sketchProcessor.helperLibraries.utils import sketch_segmentation as ss
from sketchProcessor.helperLibraries.utils import contour_classification as cc
image = cv2.imread('./testData/sketchMaps/stamps/GoodOne.bmp')
draw.showImage(image)
svg = cc.completeClassification(image, fileName='preprocessedSketchMap50')
image = cv2.imread('./testData/sketchMaps/stamps/s0.bmp')
cc.completeClassification(image, fileName='hans')
svg = cc.completeClassification(image, fileName='preprocessedSketchMap30')
def regionDetectionColor(image, filter='hsv', color='red', display=False):
# Retrieves bounding boxes that surround the image objects that have a red or a dark blue color (usually the stamps)
hsvMin = None
hsVMax = None
rgbMin = None
rgbMax = None
# Defining the number of dilation iterations and the dilation kernel
iterations = 1
kernel = np.ones((3, 3), np.uint8)
h, w, c = image.shape
if w > 6000 or h > 6000:
# If one of sides in the input image is greater than 6000 pixels
# we apply larger kernels and more iterations
iterations = 4
kernel = np.ones((5, 5), np.uint8)
if color == 'blue':
# These values were determined manually using the range-detector script in the imutils library
hsvMin = (103, 28, 144)
hsVMax = (166, 255, 255) #
rgbMin = (174, 79, 0)
rgbMax = (255, 166, 219)
else:
hsvMin = (170, 15, 0)
hsVMax = (255, 255, 255)
rgbMin = (0, 0, 189)
rgbMax = (232, 199, 255)
clone = image.copy()
mask = None
blank_image = np.zeros((image.shape[0], image.shape[1], 3))
# Here, we filter the input image, thus, creating a mask that only contains the elements whose RGV or HSV values are inside
# the stablished ranges
if (filter == 'hsv'):
clone = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
mask = cv2.inRange(clone, hsvMin, hsVMax)
mask = cv2.dilate(mask, kernel, iterations=iterations)
if display:
draw.showImage(mask)
else:
mask = cv2.inRange(clone, rgbMin, rgbMax)
mask = cv2.dilate(mask, kernel, iterations=iterations)
if display:
draw.showImage(mask)
# We find contours on the masked image.
cnts = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if imutils.is_cv2() else cnts[1]
clone = image.copy()
cnts = ss.remove_small_contours(cnts)
clone = image.copy()
#Generating bounding boxes out of the extracted contours
bBoxes2 = getContourBoundlingBox(cnts, image=image)
#Merging boxes whose overlapping area is at leat 15%
bBoxes4 = generateRegions(np.array(bBoxes2), 0.15)
# After merging some artifacts or redundant bounding boxes appear
# Here we sort the bounding boxes by area and eliminate those whose overlapping area is greater than 10%
bBoxes5 = nms.non_max_suppression_fast_area(bBoxes4, 0.1)
# Bounding boxes are padded to increase the model accuracy
boxes6 = padBoxes(image, bBoxes5, increment=10)
if display:
# Showing step by step the whole procedure
displayBoundingBoxes(image, bBoxes2, path="./", name="test")
displayBoundingBoxes(image, bBoxes4, path="./", name="test")
displayBoundingBoxes(image, bBoxes5, path="./", name="test")
displayBoundingBoxes(image, boxes6, path="./", name="test")
return boxes6