问题描述
所以我想使用 GAN 从超声图像模拟 CT 图像,我目前正在准备数据。
我相信这样更容易模拟 CT 图像。
我使用的是简单的 ITK。我想这应该是一个常见的转变。 是否有我不知道的来自 sITK 的过滤器?或者有没有其他简单的方法来进行这种转换?
解决方法
单应性的想法不起作用,所以这不会作为答案,但希望其中一些仍然有帮助。
我基本上针对六个关键点并试图纠正它们。然而,单应性没有处理顶部和底部的圆柱曲线。
import cv2
import numpy as np
# load image
img = cv2.imread("original.png");
# chop bottom (there's a weird gray band down there)
h,w = img.shape[:2];
img = img[:h-10,:,:];
# convert color
gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY);
thresh = cv2.inRange(gray,30,255);
# split
h,w = gray.shape;
half = int(w/2);
left = gray[:,:half];
right = gray[:,half:];
# find corners
threshold = 30;
# top left
stop = False;
tl = [-1,-1];
for y in range(h):
for x in range(half):
if left[y,x] > threshold:
tl = [x,y];
stop = True;
break;
if stop:
break;
# top right
stop = False;
tr = [-1,-1];
for y in range(h):
for x in range(half):
if right[y,x] > threshold:
tr = [x + half,y];
stop = True;
break;
if stop:
break;
# bottom left
bl = [-1,-1];
stop = False;
for x in range(half):
for y in range(h):
if left[y,x] > threshold:
bl = [x,y];
stop = True;
break;
if stop:
break;
# bottom right
br = [-1,-1];
stop = False;
for x in range(half - 1,-1):
for y in range(h):
if right[y,x] > threshold:
br = [x + half,y];
stop = True;
break;
if stop:
break;
# middle top
mt = [-1,-1];
for y in range(h):
if right[y,0] > threshold:
mt = [half,y];
# middle bottom
mb = [-1,-1];
for y in range(h-1,-1):
if right[y,0] > threshold:
mb = [half,y];
# corners
corners = [];
corners.append(tl);
corners.append(tr);
corners.append(br);
corners.append(bl);
corners.append(mt);
corners.append(mb);
# draw points
for p in corners:
print(p);
tup = (p[0],p[1]);
img = cv2.circle(img,tup,10,(0,255),-1);
# img = cv2.circle(img,(100,100),1000,-1);
print("Res: " + str([w,h]));
# create homography destination
targets = [];
targets.append([0,0]); # tl
targets.append([w,0]); # tr
targets.append([w,h]); # br
targets.append([0,h]); # bl
targets.append([half,0]); # mt
targets.append([half,h]); # mb
# make blank
corners = np.array(corners);
targets = np.array(targets);
hmat,ret = cv2.findHomography(corners,targets);
# warp image
warped = cv2.warpPerspective(img,hmat,(w,h));
# show
cv2.imshow("img",img);
cv2.imshow("thresh",thresh);
cv2.imshow("warped",warped);
cv2.waitKey(0);