def build_gauss_pyr(self, base): #---- 建立高斯金字塔
intvls = self.intvls
# 影像金字塔每一層有幾個分頁 =3
sigma = self.sigma
# 影像最開始的 高斯平滑參數=1.6
octvs = self.octvs
# 高斯金字塔有幾層
k = 2**(1 / intvls)
# =2^(0.333)
sig = np.zeros(intvls + 3)
# 每一頁的 sigma
sig[0] = sigma
sig[1] = sigma * np.sqrt(k * k - 1)
for i in range(2, intvls + 3):
sig[i] = sig[i - 1] * k
gauss_pyr = []
# 輸出高斯金字塔
for o in range(0, octvs):
if o == 0:
pyramid = np.zeros((base.shape[0], base.shape[1], intvls + 3))
pyramid[:, :, 0] = base
else:
pyramid = self.downsample_by2(gauss_pyr[o - 1][:, :, intvls])
for i in range(1, intvls + 3):
pyramid[:, :, i] = CVP.Gauss_blur2D(pyramid[:, :, i - 1],
sigma=sig[i])
gauss_pyr.append(pyramid)
return gauss_pyr
def pre_processing(self, src_img):
# 執行以下動作: (1)轉灰階圖 (2)將資料改成[0~1] (3)影像放大2倍 (4)Gauss smooth
gray = CVP.rgb2gray(src_img)
# 灰階圖
if np.max(gray) > 2:
gray = gray / 256
# 影像強度正規化到 [0 ~ 1]
#doubled = CVP.resize(gray,ratio=2); # 雙倍影像圖
if self.doubled == True:
doubled = CVP.resize(gray, ratio=2)
# 雙倍影像圖
else:
doubled = gray
sig_diff = np.sqrt(self.sigma**2 - 4 * self.init_sigma**2)
# s=1.25
init_img = CVP.Gauss_blur2D(doubled, sigma=sig_diff)
return init_img
def plot_pyramid(pyr, start=0, end=0, **kwargs):
"""
@ src : 影像金字塔
@ start: 從第幾層開始
@ end : 到弟幾層
@ figx,figy: 影像大小
"""
layers = np.size(pyr)
# 金字塔有幾層
if end == 0:
end = layers
# 最後第幾層
if pyr[0].ndim == 2:
pages = 1
# 每一層有幾個分頁
if pyr[0].ndim == 3: pages = pyr[0].shape[2]
ymax = 0
xmax = 0
ys = np.zeros(end - start)
# 每一層的起始位置Y
#---- ----
for octv in range(start, end): #-- 計算位置
ys[octv - start] = ymax
ymax += pyr[octv].shape[0]
xmax = np.maximum(xmax, pyr[octv].shape[1] * pages)
dest = np.zeros((ymax, xmax))
#---- ----
for octv in range(start, end): #-- 填入資料
base = pyr[octv]
rows = base.shape[0]
cols = base.shape[1]
for page in range(0, pages):
if pages > 1: img = base[:, :, page]
if pages == 1: img = base[:, :]
y1 = ymax - (ys[octv - start]).astype(int) - rows
x0 = page * cols
dest[y1:y1 + rows, x0:x0 + cols] = img
if kwargs.get('normalize'): #-- 調整動態範微
dmax = np.max(dest)
dmin = np.min(dest)
dest = (dest - dmin) / (dmax - dmin)
dmax = np.max(dest)
dmin = np.min(dest)
CVP.imshow(dest, **kwargs)
plt.show()
def focal_correct(PAN): #---- project image to cylindrical
""" imgs[k]: kth source image without correction
imgc[k]: kth image with focal length correction
"""
hfov= PAN.hfov;
for k in range(0, PAN.count):
src = PAN.imgs[k];
imgc= CVP.projection_3D(src, hfov= hfov, fl=None, mode='plane_to_cylindrical');
PAN.imgc.append(imgc);
def brightness_compensation(PAN): #---- brightness compensation
#--- reference (no boundary but blurred) ---
ref = pano_tools.paste_and_blend(PAN, src= PAN.imgb, bw=70);
gref= CVP.rgb2gray(ref);
H = PAN.H.copy();
xa = np.argmin(H[:,6]); x0 = int(np.ceil(PAN.imgb[xa].shape[1]/2));
ya = np.argmin(H[:,5]); y0 = int(np.ceil(PAN.imgb[ya].shape[0]/2));
box_ratio = 0.99;
#--- ---
for k in range(0,PAN.count):
#--- capture source sampling window ---
imgb= CVP.rgb2gray(PAN.imgb[k]);
rowh= int(imgb.shape[0]/2); # half image height
colh= int(imgb.shape[1]/2); # half image width
roww= int(rowh* box_ratio); # capture window radius
colw= int(colh* box_ratio);
imgs= imgb[rowh-roww:rowh+roww, colh-colw:colh+colw]
cdfs= pano_tools.brightness_cdf(imgs); # brightness curve for source
#--- capture reference sampling window ---
yr = int(y0+ H[k,5]);
xr = int(x0+ H[k,6]);
imgr= gref[yr-roww:yr+roww, xr-colw:xr+colw];
cdfr= pano_tools.brightness_cdf(imgr); # brightness curve for reference
#--- brightness equalization ---
out = np.zeros_like(PAN.imgb[k]); # image with brightness compensation
for kb in range(0,256):
under= np.sum(cdfr< cdfs[kb]); #
under= np.minimum(under, 3*(kb+0.5)-0.5); # prevent from low gain boost
gain = (under+0.5)/(kb+0.5);
#--- find pixel---
th0 = kb/256; th1 = (kb+1)/256; # bin lower and upper bound
mask = ((imgb<th1)+0) - ((imgb<=th0)+0) # data inside the region
mask = mask* gain;
out[:,:,0]+= PAN.imgt[k][:,:,0]*mask;
out[:,:,1]+= PAN.imgt[k][:,:,1]*mask;
out[:,:,2]+= PAN.imgt[k][:,:,2]*mask;
PAN.imgb[k] = out;
PAN.upload(plt.imread('taipei_03.jpg') / 256)
PAN.upload(plt.imread('taipei_04.jpg') / 256)
PAN.upload(plt.imread('taipei_05.jpg') / 256)
pano_tools.plot_source_images(PAN)
# plot source images
#=================================================== Auto execution
if 1 == 0:
pano_tools.focal_correct(PAN)
pano_tools.extract_features(PAN)
# extract features
pano_tools.pairing(PAN, th=None)
# pairing feature points
pano_tools.floorplan(PAN, tilt1=3.5)
pano_tools.transform(PAN)
pano_tools.registry(PAN, bw=25, box_th=None)
CVP.imshow(PAN.result, figy=8, ticks='off')
#=================================================== Step 1: focal length correct
if 1 == 0:
pano_tools.focal_correct(PAN)
pano_tools.plot_focal_correct(PAN)
# plot source images
#=================================================== Step 2: extract features
if 1 == 0:
pano_tools.extract_features(PAN)
# extract features
pano_tools.plot_features(PAN)
# plot keypoints for image 0
#=================================================== Step 3: Pairing
if 1 == 0:
pano_tools.pairing(PAN, th=None)
def plot_transform(PAN): #---- plot registry images
for k in range(0, np.shape(PAN.imgt)[0]):
CVP.imshow(PAN.imgt[k],figy=6,ticks='off');
print('image no:',k);
def plot_focal_correct(PAN): #---- plot source image base
for k in range(0, np.shape(PAN.imgc)[0]):
CVP.imshow(PAN.imgc[k],figy=6,ticks='off');
print('image no:',k);
def plot_source_images(PAN) : #---- plot source image base
for k in range(0, np.shape(PAN.imgs)[0]):
CVP.imshow(PAN.imgs[k],figy=6,ticks='off');
print('image no:',k);