def hierEncode(self, f_half, f): highResShape = np.shape(f) E_f_half = imt.ImageTransform.imresize(self.IMT.rgb2img(f_half), highResShape, 1) E_f_half_RGB = np.array(E_f_half) diffRGB = f.astype(int) - E_f_half_RGB.astype(int) diffRGB[diffRGB<0] = 1 diffRGB[diffRGB>255] = 255 diffRGB = diffRGB.astype(np.uint8) F_info, orgShape = coding.encode(np.array(diffRGB), self.quantMat) recRGB, recImg2show = coding.decode(F_info, self.quantMat, orgShape) return F_info, orgShape, recRGB, recImg2show
def hierEncode(self, f_half, f):
highResShape = np.shape(f)
E_f_half = imt.ImageTransform.imresize(self.IMT.rgb2img(f_half),
highResShape, 1)
E_f_half_RGB = np.array(E_f_half)
diffRGB = f.astype(int) - E_f_half_RGB.astype(int)
diffRGB[diffRGB < 0] = 1
diffRGB[diffRGB > 255] = 255
diffRGB = diffRGB.astype(np.uint8)
F_info, orgShape = coding.encode(np.array(diffRGB), self.quantMat)
recRGB, recImg2show = coding.decode(F_info, self.quantMat, orgShape)
return F_info, orgShape, recRGB, recImg2show
def main(quantMat, path):
#quantMat = np.full([8,8], 1)
#Read Image
orgImage = imt.ImageTransform()
orgRGB = orgImage.readImage(path)
#Reduce resolution
imgPIL = orgImage.rgb2img(orgRGB)
halfImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 2)
quartImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 4)
'''Hierachical Encoding'''
#Encode quartImg
F_info_quart, orgShape_quart = coding.encode(np.array(quartImg), quantMat)
#print orgShape_quart
#Decode QuartImg
recRGB_quart, recImg2show_quart = coding.decode(F_info_quart, quantMat,
orgShape_quart)
#recImg2show_quart.show()
#Encode half diff Image
hierVar = hr.Hier(quantMat)
halfdiffF_info, halfdifforgShape, hafldiffrecRGB, halfdiffrecImg2show = hierVar.hierEncode(
np.array(quartImg), np.array(halfImg))
#Decode half
recRGB_halfdiff, recImg2show_halfdiff = coding.decode(
halfdiffF_info, quantMat, halfdifforgShape)
fRecoverHalf, fRecoverHalfImg = hierVar.hireDecode(np.array(quartImg),
recRGB_halfdiff)
#fRecoverHalfImg.show()
#fRecover, fRecoverImg = hierVar.hireDecode(fRecoverHalf, diffF_info, difforgShape)
#Encode whole diff Image
#imgPIL.show()
diffF_info, difforgShape, diffrecRGB, diffrecImg2show = hierVar.hierEncode(
np.array(fRecoverHalfImg), np.array(imgPIL))
#Decode whole image
recRGB_diff, recImg2show_diff = coding.decode(diffF_info, quantMat,
difforgShape)
fRecover, fRecoverImg = hierVar.hireDecode(np.array(fRecoverHalfImg),
recRGB_diff)
#fRecoverHalfImg.show()
#fRecoverImg.show()
return [
imgPIL, fRecoverImg, halfdiffrecImg2show, diffrecImg2show,
fRecoverHalfImg, recImg2show_quart
]
def main(quantMat, path): #quantMat = np.full([8,8], 1) #Read Image orgImage = imt.ImageTransform() orgRGB = orgImage.readImage(path) #Reduce resolution imgPIL = orgImage.rgb2img(orgRGB) halfImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 2) quartImg = imt.ImageTransform.imresize(imgPIL, np.shape(orgRGB), 4) '''Hierachical Encoding''' #Encode quartImg F_info_quart, orgShape_quart = coding.encode(np.array(quartImg), quantMat) #print orgShape_quart #Decode QuartImg recRGB_quart, recImg2show_quart = coding.decode(F_info_quart, quantMat, orgShape_quart) #recImg2show_quart.show() #Encode half diff Image hierVar = hr.Hier(quantMat) halfdiffF_info, halfdifforgShape, hafldiffrecRGB, halfdiffrecImg2show = hierVar.hierEncode(np.array(quartImg), np.array(halfImg)) #Decode half recRGB_halfdiff, recImg2show_halfdiff = coding.decode(halfdiffF_info, quantMat, halfdifforgShape) fRecoverHalf, fRecoverHalfImg = hierVar.hireDecode(np.array(quartImg), recRGB_halfdiff) #fRecoverHalfImg.show() #fRecover, fRecoverImg = hierVar.hireDecode(fRecoverHalf, diffF_info, difforgShape) #Encode whole diff Image #imgPIL.show() diffF_info, difforgShape, diffrecRGB, diffrecImg2show = hierVar.hierEncode(np.array(fRecoverHalfImg), np.array(imgPIL)) #Decode whole image recRGB_diff, recImg2show_diff = coding.decode(diffF_info, quantMat, difforgShape) fRecover, fRecoverImg = hierVar.hireDecode(np.array(fRecoverHalfImg), recRGB_diff) #fRecoverHalfImg.show() #fRecoverImg.show() return [imgPIL, fRecoverImg, halfdiffrecImg2show ,diffrecImg2show, fRecoverHalfImg ,recImg2show_quart]
class DatasetIAMSplit:
"""wrapper which provides a dataset interface for a split of the original dataset"""
def __init__(self, dataset, start_idx, end_idx):
assert start_idx >= 0 and end_idx <= len(dataset)
self.dataset = dataset
self.start_idx = start_idx
self.end_idx = end_idx
def __getitem__(self, idx):
return self.dataset[self.start_idx + idx]
def __len__(self):
return self.end_idx - self.start_idx
if __name__ == '__main__':
from visualization import visualize
from coding import encode, decode
import matplotlib.pyplot as plt
dataset = DatasetIAM(Path('../data'), (350, 350), (350, 350), caching=False)
img, gt = dataset[0]
gt_map = encode(img.shape, gt)
gt = decode(gt_map)
plt.imshow(visualize(img / 255 - 0.5, gt))
plt.show()
def check(s):
c = encode(s)
s2 = decode(c)
assert '\n' not in c[:-1]
assert c[-1] == '\n'
assert s == s2
def __getitem__(self, item):
batch_imgs = []
batch_gt_maps = []
batch_aabbs = []
for b in range(self.batch_size):
if self.is_random:
shuffled_idx = self.shuffled_indices[item * self.batch_size +
b]
else:
shuffled_idx = item * self.batch_size + b
img, aabbs = self.dataset[shuffled_idx]
if self.is_random:
# geometric data augmentation (image [0..255] and gt)
if prob_true(0.75):
# random scale
fx = np.random.uniform(0.5, 1.5)
fy = np.random.uniform(0.5, 1.5)
# random position around center
txc = self.input_size[1] * (1 - fx) / 2
tyc = self.input_size[0] * (1 - fy) / 2
freedom_x = self.input_size[1] // 10
freedom_y = self.input_size[0] // 10
tx = txc + np.random.randint(-freedom_x, freedom_x)
ty = tyc + np.random.randint(-freedom_y, freedom_y)
# map image into target image
M = np.float32([[fx, 0, tx], [0, fy, ty]])
white_bg = np.ones(self.input_size, np.uint8) * 255
img = cv2.warpAffine(img,
M,
dsize=self.input_size[::-1],
dst=white_bg,
borderMode=cv2.BORDER_TRANSPARENT)
# apply the same transformations to gt, and clip/remove aabbs outside of target image
aabb_clip = AABB(0, img.shape[1], 0, img.shape[0])
aabbs = [
aabb.scale(fx, fy).translate(tx, ty).clip(aabb_clip)
for aabb in aabbs
]
aabbs = [aabb for aabb in aabbs if aabb.area() > 0]
# photometric data augmentation (image [-0.5..0.5] only)
img = (img / 255 - 0.5)
if prob_true(0.25): # random distractors (lines)
num_lines = np.random.randint(1, 20)
for _ in range(num_lines):
rand_pt = lambda: (np.random.randint(0, img.shape[1]),
np.random.randint(0, img.shape[0]))
color = np.random.triangular(-0.5, 0, 0.5)
thickness = np.random.randint(1, 3)
cv2.line(img, rand_pt(), rand_pt(), color, thickness)
if prob_true(0.75): # random contrast
img = (img - img.min()) / (img.max() -
img.min()) - 0.5 # stretch
img = img * np.random.triangular(0.1, 0.9,
1) # reduce contrast
if prob_true(0.25): # random noise
img = img + np.random.uniform(-0.1, 0.1, size=img.shape)
if prob_true(0.25): # change thickness of text
img = cv2.erode(img, np.ones((3, 3)))
if prob_true(0.25): # change thickness of text
img = cv2.dilate(img, np.ones((3, 3)))
if prob_true(0.25): # invert image
img = 0.5 - img
else:
img = (img / 255 - 0.5)
gt_map = encode(self.output_size, aabbs, self.scale_down)
batch_imgs.append(img[None, ...].astype(np.float32))
batch_gt_maps.append(gt_map)
batch_aabbs.append(aabbs)
batch_imgs = np.stack(batch_imgs, axis=0)
batch_gt_maps = np.stack(batch_gt_maps, axis=0)
batch_imgs = torch.from_numpy(batch_imgs).to('cuda')
batch_gt_maps = torch.from_numpy(batch_gt_maps.astype(
np.float32)).to('cuda')
return DataLoaderItem(batch_imgs, batch_gt_maps, batch_aabbs)
def on_encode_button_clicked(self):
method = self.method_chooser.get()
text = self.text.get('1.0', 'end')[:-1]
code = coding.encode(method, text)
self.code.delete('1.0', 'end')
self.code.insert('1.0', code)
def hamming_code(self):
return encode(self.id_as_binary)
def dumps(obj, protocol=1):
return encode(pickle.dumps(obj, protocol))
