Cluster similar shapes in an image using hu moments, with the help of opencv and scikit-learn

README.md

Sample Output

Similar shapes in the input image(left) are grouped into same clusters, using hu moments. Members of the same cluster are shown with same coloured bounding boxes in the output image(right).

In the above image there are four different clusters corresponding to four shapes: Square, Rectangle, Triangle and Ellipse

contour_humom.py

import numpy as np
import cv2

from scipy.cluster.hierarchy import linkage, fcluster
from scipy.spatial import distance as ssd

# Read input image and convert to grayscale
img = cv2.imread('shapes.jpg')
print("Image shape: " + str(img.shape)) 
imgray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)


# Apply adaptive threshold
cv2.imwrite('gray.png',imgray)
thresh = cv2.adaptiveThreshold(imgray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,cv2.THRESH_BINARY_INV,11,2)
cv2.imwrite('edge.png',thresh)


# Find all contours of binary image
_, contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

# Number of contours
num_contours=len(contours)
print("Total number of contours: " + str(num_contours))

# Fill up all the shapes
filled=cv2.fillPoly(thresh, pts=contours, color=(255,255,255))

# Remove noise by morphological opening
kernel = np.ones((4,1),np.uint8)
filled=cv2.morphologyEx(filled, cv2.MORPH_OPEN, kernel)

# Crop and save proper contours separately [filter by area]
contour_crop=[]
colour_crop=[]
num_contours=0
min_area=50
for cnt in contours:
  if(cv2.contourArea(cnt)>min_area):
    x,y,w,h = cv2.boundingRect(cnt)
    contour_crop.append(filled[y:y+h, x:x+w])
    colour_crop.append(img[y:y+h, x:x+w])
    cv2.imwrite("contour_binary"+str(num_contours)+".png",filled[y:y+h, x:x+w])
    num_contours=num_contours+1

# Number of proper contours
print("Number of proper contours: " + str(num_contours))

# Initialize similarity matrix
sim_mat=np.zeros((num_contours, num_contours))

# Configure cluster settings
cluster=np.zeros(num_contours, dtype=int)
num_classes=4
colours=[(255,0,0),(0,255,0),(0,0,255),(0,0,0),(255,255,0), (0,255,255),(255,0,255),(128,128,128),(165,42,42)]


# Compute similarity matrix [Uses Hu-Moments]
print("Similarity Matrix - Hu Moements")
for i in range(0,num_contours):
   for j in range(0, num_contours):
      im1=np.array(contour_crop[i])
      im2=np.array(contour_crop[j])
      sim_mat[i][j]=cv2.matchShapes(im1,im2,cv2.CONTOURS_MATCH_I2,0)
      print(str(round(sim_mat[i][j],4))+",",end =" ")
   print("\n")


'''
 Cluster the contours using similarity matrix and max cluster no.
 Returns cluster vector  [Range: 0 - (num_classes-1)]

'''
Zd = linkage(ssd.squareform(sim_mat), method="complete") 
cluster = fcluster(Zd, num_classes, criterion='maxclust') - 1  
print("Cluster:-\n" + str(cluster))


# Draw contour clusters
c=0
for cnt in contours:
  if(cv2.contourArea(cnt)>min_area):
    x,y,w,h = cv2.boundingRect(cnt)
    cv2.imwrite("contour_coloured"+str(c)+".png",img[y:y+h, x:x+w])
    cv2.rectangle(img,(x,y),(x+w,y+h),colours[cluster[c]],2)
    c=c+1           

# Show output in window
cv2.imshow("Bounding Boxes", img)
cv2.waitKey(0)
cv2.destroyAllWindows()