shape_matching/match.py

168 lines
6.1 KiB
Python
Raw Normal View History

2024-08-11 16:54:51 +02:00
# import required libraries
import cv2
import numpy as np
import matplotlib
import math
MAX = 10 # should currently be 10
NormUpperBound = 2.0 #Highest Expected Number
GrayscaleThreshhold = 200
RotationAngle = 45
### Read Images from Disk, Assumed Naming: 0.png 1.png ... n.png depending on MAX
#Turns Images into Contours based on Greyscale Threshhold
def GetContoursFromFiles():
cnt = []
for i in range (0, MAX, 1):
# Read image as grayscale images
img = cv2.imread(str(i)+'.png',0)
# Apply thresholding on the images to convert to binary images
ret, thresh1 = cv2.threshold(img, GrayscaleThreshhold, 255,cv2.THRESH_BINARY_INV)
# find the contours in the binary image
contours,hierarchy = cv2.findContours(thresh1,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
print("Number of Shapes detected in Image "+str(i)+":",len(contours))
cnt.append(contours[0])
return cnt
#Does N to N matching, resulting in a 1d array that should be interpreted as a 2d array
# Could easily be modified for 1 to N matching
def MatchAll(cnt):
mat = []
for i in range(0, MAX, 1):
for j in range(0, MAX, 1):
mat.append(cv2.matchShapes(cnt[i],cnt[j],1,0.0))
return mat
def PrintMatchingMatrix(mat):
print("Similarity of Images: \n")
for i in range(0, MAX, 1):
for j in range(0, MAX, 1):
print(f"{mat[i*MAX+j]:.2f}", end='')
print("\t", end='')
print("\n")
def PrintMatchingMatrixNormalized(mat):
print("Similarity of Images Normalized: \n")
norm = matplotlib.colors.Normalize(vmin=0.0, vmax=NormUpperBound)
for i in range(0, MAX, 1):
for j in range(0, MAX, 1):
print(f"{norm(mat[i*MAX+j]):.2f}", end='')
print("\t", end='')
print("\n")
# Does all of the heavy lifting when it comes to displaying it in a nice graphical way
# Compact but also quickly hacked together
# Builds the Visual matrix using images, 1 image = 1 coordinate unit
def CreateMatchingColorMatrix(mat):
im_res = cv2.imread(str(0)+'.png',cv2.IMREAD_COLOR)
height, width, channels = im_res.shape
# (Coordinate 0/0) Color (white)
im_temp = np.full((height, width, 3), 255, np.uint8)
norm = matplotlib.colors.Normalize(vmin=0.0, vmax=NormUpperBound)
# Build Topmost row (just iterate through all images and concat them sequentially)
for i in range(0, MAX, 1):
img = cv2.imread(str(i)+'.png',cv2.IMREAD_COLOR)
im_temp = cv2.hconcat([im_temp, img])
# This top row is now our first row
im_res = im_temp
# Build The matrix row by row
for i in range(0, MAX, 1):
im_temp = cv2.imread(str(i)+'.png',cv2.IMREAD_COLOR)
img = np.full((height, width, 3), 255, np.uint8)
img[:] = (0, 0, 255)
# Individual row here, current sequential image gets chosen above, so here, we can do square coloring
for j in range(0, MAX, 1):
cmap = matplotlib.cm.get_cmap('brg_r')
cmap.set_over((0.0,0.0,0.0))
# Gets current weight, normalises it, looks it up in color map, converts it to full scale, colors
img[:] = NtoF(cmap(norm(mat[i*MAX+j]))[:-1])
# build up row
im_temp = cv2.hconcat([im_temp, img])
#build up matrix
im_res = cv2.vconcat([im_res, im_temp])
DebugDrawImage(im_res)
# Helper to convert Normalized color to Full-scale (255) color
def NtoF(rgb):
return tuple([255*x for x in rgb])
def DebugDrawImage(img):
cv2.imshow("Image", img)
cv2.waitKey(0)
### Rotates Image 0.png 4 times to test how rotations affect outcome (they don't)
def RotationTest():
global MAX
MAX = math.ceil(360/RotationAngle)
cnt = []
img = cv2.imread(str(0)+'.png',0)
for i in range (0, MAX, 1):
ret, thresh1 = cv2.threshold(img, GrayscaleThreshhold, 255,cv2.THRESH_BINARY_INV)
contours,hierarchy = cv2.findContours(thresh1,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
cnt.append(contours[0])
print("Number of Shapes detected in Rotation "+str(i)+":",len(contours))
img = rotate_image(img, RotationAngle)
mat = MatchAll(cnt)
PrintMatchingMatrix(mat)
CreateMatchingColorMatrixRotation(mat)
# Quick hack of above function to work with rotation instead of multiple images
def CreateMatchingColorMatrixRotation(mat):
im_rot = cv2.imread(str(0)+'.png',cv2.IMREAD_COLOR)
im_res = cv2.imread(str(0)+'.png',cv2.IMREAD_COLOR)
height, width, channels = im_res.shape
im_temp = np.full((height, width, 3), 255, np.uint8)
norm = matplotlib.colors.Normalize(vmin=0.0, vmax=NormUpperBound)
for i in range(0, MAX, 1):
img = im_rot
im_temp = cv2.hconcat([im_temp, img])
im_rot = rotate_image(img, RotationAngle)
im_res = im_temp
#reset
im_rot = cv2.imread(str(0)+'.png',cv2.IMREAD_COLOR)
for i in range(0, MAX, 1):
im_temp = cv2.imread(str(0)+'.png',cv2.IMREAD_COLOR)
im_temp = im_rot
img = np.full((height, width, 3), 255, np.uint8)
img[:] = (0, 0, 255)
for j in range(0, MAX, 1):
cmap = matplotlib.cm.get_cmap('brg_r')
cmap.set_over((0.0,0.0,0.0))
img[:] = NtoF(cmap(norm(mat[i*MAX+j]))[:-1])
im_temp = cv2.hconcat([im_temp, img])
im_res = cv2.vconcat([im_res, im_temp])
im_rot = rotate_image(im_rot, RotationAngle)
DebugDrawImage(im_res)
# Stolen from Stackoverflow and modified: https://stackoverflow.com/a/9042907
# NOTE: this function assumes the background is a lighter shade than the form to detect
# NOTE: INTER_NEAREST or other interpolation strategies might work better, depending on rotation and some other factors
def rotate_image(image, angle):
image_center = tuple(np.array(image.shape[1::-1]) / 2)
rot_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)
result = cv2.warpAffine(image, rot_mat, image.shape[1::-1], flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT, borderValue=(255,255,255))
return result
if __name__ == "__main__":
mat = MatchAll(GetContoursFromFiles())
PrintMatchingMatrix(mat)
PrintMatchingMatrixNormalized(mat)
CreateMatchingColorMatrix(mat)
#RotationTest()