def get_target_image(parsed, target, path=None, file=None, raw_stitch=False):
if PAIR in target:
(a,b) = target[PAIR]
a_image = read_target_image(a, path=path, file=file)
b_image = read_target_image(b, path=path, file=file)
if raw_stitch:
return stitch_raw((a,b),(a_image,b_image),background=180)
else:
try:
bin_zip = get_product_file(parsed, BINZIP_PRODUCT)
if os.path.exists(bin_zip):
# name is target LID + png extension
png_name = os.path.basename(target[PID]) + '.png'
png_data = get_zip_entry_bytes(bin_zip, png_name)
pil_img = PIL.Image.open(StringIO(png_data))
return np.array(pil_img.convert('L')) # convert to 8-bit grayscale
except NotFound:
pass
im,_ = v1_stitching.stitch((a,b),(a_image,b_image))
return im
else:
return read_target_image(target, path=path, file=file)
def stitch(targets,images):
# compute bounds relative to the camera field
(x,y,w,h) = stitched_box(targets)
# note that w and h are switched from here on out to rotate 90 degrees.
# step 1: compute masks
s = as_pil(stitch_raw(targets,images,(x,y,w,h))) # stitched ROI's with black gaps
rois_mask = as_pil(mask(targets)) # a mask of where the ROI's are
gaps_mask = ImageChops.invert(rois_mask) # its inverse is where the gaps are
edges = edges_mask(targets,images) # edges are pixels along the ROI edges
# step 2: estimate background from edges
# compute the mean and variance of the edges
(mean,variance) = bright_mv(s,edges)
# now use that as an estimated background
flat_bg = Image.new('L',(h,w),mean) # FIXME
s.paste(mean,None,gaps_mask)
# step 3: compute "probable background": low luminance delta from estimated bg
bg = extract_background(s,flat_bg)
# also mask out the gaps, which are not "probable background"
bg.paste(255,None,gaps_mask)
# step 3a: improve mean/variance estimate
(mean,variance) = bright_mv(bg)
std_dev = sqrt(variance)
# step 4: sample probable background to compute RBF for illumination gradient
# grid
div = 6
means = []
nodes = []
rad = avg([h,w]) / div
rad_step = int(rad/2)+1
for x in range(0,h+rad,rad):
for y in range(0,w+rad,rad):
for r in range(rad,max(h,w),int(rad/3)+1):
box = (max(0,x-r),max(0,y-r),min(h-1,x+r),min(w-1,y+r))
region = bg.crop(box)
nabe = region.histogram()
(m,v) = bright_mv_hist(nabe)
if m > 0 and m < 255: # reject outliers
nodes.append((x,y))
means.append(m)
break
# now construct radial basis functions for mean, based on the samples
mean_rbf = interpolate.Rbf([x for x,y in nodes], [y for x,y in nodes], means, epsilon=rad)
# step 5: fill gaps with mean based on RBF and variance from bright_mv(edges)
mask_pix = gaps_mask.load()
noise = Image.new('L',(h,w),mean)
noise_pix = noise.load()
np.random.seed(0)
gaussian = np.random.normal(0, 1.0, size=(h,w)) # it's normal
std_dev *= 0.66 # err on the side of smoother rather than noisier
mask_x = []
mask_y = []
for x in xrange(h):
for y in xrange(w):
if mask_pix[x,y] == 255: # only for pixels in the mask
mask_x.append(x)
mask_y.append(y)
rbf_fill = mean_rbf(np.array(mask_x), np.array(mask_y))
for x,y,r in zip(mask_x, mask_y, rbf_fill):
# fill is illumination gradient + noise
noise_pix[x,y] = r + (gaussian[x,y] * std_dev)
# step 6: final composite
s.paste(noise,None,gaps_mask)
return (np.array(s),rois_mask)
