I=io.imread(os.path.join(dataDir,dataType,'{:012d}.jpg'.format(img['id'])))
annIds = coco.getAnnIds(imgIds=img['id'], catIds=catIds, iscrowd=None)
anns = coco.loadAnns(annIds)
bbx=anns[0]['bbox']
mask=np.array(coco.annToMask(anns[0]))
print(np.shape(mask))
print(np.shape(I))
I1=I
# mask 是 01 蒙版
I1[:,:,0]=np.array(I[:,:,0] * mask )
I1[:,:,1]=np.array(I[:,:,1] * mask )
I1[:,:,2]=np.array(I[:,:,2] * mask )
# differece_8是background的edge
difference_8 = mask_to_outeedge(mask)
# check the outer edge is right
# feedback1: generate difference_8 wrong
check_outer_edge = np.where(difference_8 == 1,0.5,0)+mask
# plt.figure('check_outer_edge')
# plt.imshow(check_outer_edge)
# plt.show()
difference_8_dilation = dilation.binary_dilation(difference_8, np.ones((3,3)))
difference_8_dilation = np.where(difference_8_dilation == True,1,0)
# double_edge_candidate = difference_8_dilation - difference_8
double_edge_candidate = difference_8_dilation + mask
double_edge = np.where(double_edge_candidate == 2, 1, 0) # here we get double edge,now to confirm it
confirm_image = np.where(double_edge==1,0.5,0) + np.where(difference_8 == 1,0.8,0)
ground_truth = np.where(double_edge==1, 255, 0) + np.where(difference_8 == 1, 100, 0) + np.where(mask == 1, 50 , 0) # 所以内侧边缘就是100的灰度值
def splicing_tamper_one_image(input_foreground,
input_background,
mask,
similar_judge=True):
"""
在输入这个函数之前，输入的两个参数需要以及经过挑选了的,需要import poisson融合的代码
:param foreground:
:param background:
:return: 返回两个参数：直接篡改图, poisson融合篡改图, GT
"""
I = input_foreground
I1 = I
# mask 是 01 蒙版
I1[:, :, 0] = np.array(I[:, :, 0] * mask)
I1[:, :, 1] = np.array(I[:, :, 1] * mask)
I1[:, :, 2] = np.array(I[:, :, 2] * mask)
# differece_8是background的edge
difference_8 = mask_to_outeedge(mask)
difference_8_dilation = dilation.binary_dilation(difference_8,
np.ones((3, 3)))
difference_8_dilation = np.where(difference_8_dilation == True, 1, 0)
double_edge_candidate = difference_8_dilation + mask
double_edge = np.where(double_edge_candidate == 2, 1, 0)
ground_truth = np.where(double_edge == 1, 255, 0) + np.where(
difference_8 == 1, 100, 0) + np.where(mask == 1, 50,
0) # 所以内侧边缘就是100的灰度值
b1 = input_background
background = Image.fromarray(b1, 'RGB')
foreground = Image.fromarray(I1, 'RGB').convert('RGBA')
datas = foreground.getdata()
newData = []
for item in datas:
if item[0] == 0 and item[1] == 0 and item[2] == 0:
newData.append((0, 0, 0, 0))
else:
newData.append(item)
foreground.putdata(newData)
background = background.resize((foreground.size[0], foreground.size[1]),
Image.ANTIALIAS)
# input_background = np.resize(input_background,(foreground.size[1], foreground.size[0],3))
background_area = np.array(background)
try:
if similar_judge:
foreground_area = I1
background_area[:, :,
0] = np.array(background_area[:, :, 0] * mask)
background_area[:, :,
1] = np.array(background_area[:, :, 1] * mask)
background_area[:, :,
2] = np.array(background_area[:, :, 2] * mask)
foreground_area = Image.fromarray(foreground_area)
background_area = Image.fromarray(background_area)
if myCalImage.calc_similar(foreground_area, background_area) > 0.2:
pass
else:
return False, False, False
except Exception as e:
print(e)
traceback.print_exc()
try:
mask = Image.fromarray(mask)
except Exception as e:
print('mask to Image error', e)
# 在这里的时候，mask foreground background 尺寸都是一致的了，poisson融合时，offset置为0
try:
poisson_foreground = cv2.cvtColor(
np.asarray(foreground.convert('RGB')), cv2.COLOR_RGB2BGR)
poisson_background = cv2.cvtColor(np.asarray(background),
cv2.COLOR_RGB2BGR)
poisson_mask = np.asarray(mask)
poisson_mask = np.where(poisson_mask == 1, 255, 0)
poisson_fusion_image = poisson_image_editing.poisson_fusion(
poisson_foreground, poisson_background, poisson_mask)
poisson_fusion_image = Image.fromarray(
cv2.cvtColor(poisson_fusion_image, cv2.COLOR_BGR2RGB))
background.paste(foreground, (0, 0), mask=foreground.split()[3])
return background, poisson_fusion_image, ground_truth
except Exception as e:
traceback.print_exc()
def specific_object_to_fix_background(background,
mask,
tamper_num=2,
bk_shape=(320, 320)):
"""
输入一张背景图和一张mask图，该mask图的size不应需要与背景图一致
:param background:
:param mask: mask任意一个前景图的mask
:param tamper_num :篡改的数量，默认为1
:return: 篡改好了的图像和一张GT 是list
"""
background = image_crop.crop(background, target_shape=(320, 320))
try:
if background == 'error':
return False, False, False
except Exception as e:
traceback.print_exc()
sys.exit()
# 找到mask 的矩形区域
oringal_background = background.copy()
a = mask
a = np.where(a != 0)
bbox = np.min(a[0]), np.min(a[1]), np.max(a[0]), np.max(a[1])
cut_mask = mask[bbox[0]:bbox[2], bbox[1]:bbox[3]]
# 以左上角的点作为参考点，计算可以paste的区域
background_shape = background.shape
object_area_shape = cut_mask.shape
paste_area = [
background_shape[0] - object_area_shape[0],
background_shape[1] - object_area_shape[1]
]
print('the permit paste area is :', paste_area)
row1 = np.random.randint(0, paste_area[0])
col1 = np.random.randint(0, paste_area[1])
# 在background上获取mask的区域
temp_background = background.copy()
cut_area = temp_background[row1:row1 + object_area_shape[0],
col1:col1 + object_area_shape[1], :]
cut_area[:, :, 0] = cut_area[:, :, 0] * cut_mask
cut_area[:, :, 1] = cut_area[:, :, 1] * cut_mask
cut_area[:, :, 2] = cut_area[:, :, 2] * cut_mask
for i in range(3):
row2 = np.random.randint(0, paste_area[0])
col2 = np.random.randint(0, paste_area[1])
if abs(row1 - row2) + abs(col1 - col2) < 10:
print('随机选到的区域太近，最好重新选择')
else:
break
# 判断object和bg的大小是否符合要求
if paste_area[0] < 5 or paste_area[1] < 5:
print('提醒：允许的粘贴区域太小')
if paste_area[0] < 1 or paste_area[1] < 1:
print('无允许粘贴的区域')
return False, False, False
# 随机在background上贴上该mask的区域，并且保证与原区域有一定的像素偏移,然后生成新的mask图
tamper_image = []
tamper_mask = []
tamper_gt = []
tamper_poisson = []
for times in range(tamper_num):
bk_mask = np.zeros((background_shape[0], background_shape[1]),
dtype='uint8')
bk_area = np.zeros((background_shape[0], background_shape[1], 3),
dtype='uint8')
bk_mask[row2:row2 + object_area_shape[0],
col2:col2 + object_area_shape[1]] = cut_mask
bk_area[row2:row2 + object_area_shape[0],
col2:col2 + object_area_shape[1], :] = cut_area
background[:, :,
0] = background[:, :, 0] * np.where(bk_mask == 1, 0, 1)
background[:, :,
1] = background[:, :, 1] * np.where(bk_mask == 1, 0, 1)
background[:, :,
2] = background[:, :, 2] * np.where(bk_mask == 1, 0, 1)
background = background + bk_area
tamper_image.append(background)
tamper_mask.append(bk_mask)
# 调用save_method保存
for index, item in enumerate(tamper_image):
difference_8 = mask_to_outeedge(tamper_mask[index])
difference_8_dilation = dilation.binary_dilation(
difference_8, np.ones((3, 3)))
difference_8_dilation = np.where(difference_8_dilation == True, 1, 0)
double_edge_candidate = difference_8_dilation + tamper_mask[index]
double_edge = np.where(double_edge_candidate == 2, 1, 0)
ground_truth = np.where(double_edge == 1, 255, 0) + np.where(
difference_8 == 1, 100, 0) + np.where(tamper_mask[index] == 1, 50,
0) # 所以内侧边缘就是100的灰度值
tamper_gt.append(ground_truth)
try:
mask = Image.fromarray(tamper_mask[index])
except Exception as e:
print('mask to Image error', e)
# 在这里的时候，mask foreground background 尺寸都是一致的了，poisson融合时，offset置为0
foreground = Image.fromarray(bk_area)
background = Image.fromarray(oringal_background)
mask = bk_mask
try:
poisson_foreground = cv2.cvtColor(
np.asarray(foreground.convert('RGB')), cv2.COLOR_RGB2BGR)
poisson_background = cv2.cvtColor(np.asarray(background),
cv2.COLOR_RGB2BGR)
poisson_mask = np.asarray(mask)
poisson_mask = np.where(poisson_mask == 1, 255, 0)
poisson_fusion_image = poisson_image_editing.poisson_fusion(
poisson_foreground, poisson_background, poisson_mask)
poisson_fusion_image = Image.fromarray(
cv2.cvtColor(poisson_fusion_image, cv2.COLOR_BGR2RGB))
#
# plt.figure('123')
# plt.imshow(poisson_fusion_image)
# plt.show()
tamper_poisson.append(poisson_fusion_image)
return tamper_image, tamper_poisson, tamper_gt
except Exception as e:
traceback.print_exc()
def paste_object_to_background(foreground,
background,
mask,
bbox,
tamper_num=1,
optimize=False):
"""
拿一张前景图 和mask 图 将mask区域paste到background上
:param foreground:
:param background:
:param mask:
:param tamper_num: return 多少张篡改图——位置不一样 ，默认为1
:param optimize: 是否使用优化算法，寻找最佳区域
:return:
"""
# 把特定区域的object给弄出来`
orignal_bg = background
object_area = foreground
object_area[:, :, 0] = np.array(object_area[:, :, 0] * mask)
object_area[:, :, 1] = np.array(object_area[:, :, 1] * mask)
object_area[:, :, 2] = np.array(object_area[:, :, 2] * mask)
a = mask
a = np.where(a != 0)
bbox = np.min(a[0]), np.min(a[1]), np.max(a[0]), np.max(a[1])
cut_mask = mask[bbox[0]:bbox[2], bbox[1]:bbox[3]]
object_area = object_area[bbox[0]:bbox[2], bbox[1]:bbox[3], :]
#
# plt.figure('获取矩形区域')
# plt.imshow(object_area)
# plt.show()
#
# 以左上角的点作为参考点，计算可以paste的区域
background_shape = background.shape
object_area_shape = cut_mask.shape
paste_area = [
background_shape[0] - object_area_shape[0],
background_shape[1] - object_area_shape[1]
]
print('the permit paste area is :', paste_area)
row = np.random.randint(0, paste_area[0])
col = np.random.randint(0, paste_area[1])
# 判断object和bg的大小是否符合要求
if paste_area[0] < 5 or paste_area[1] < 5:
print('提醒：允许的粘贴区域太小')
tamper_image = []
tamper_poisson = []
tamper_mask = []
tamper_gt = []
if optimize == False:
for times in range(tamper_num):
try:
padding_mask = np.zeros(
(background_shape[0], background_shape[1]))
padding_mask[row:row + object_area_shape[0],
col:col + object_area_shape[1]] = cut_mask
padding_object_area = np.zeros(
(background_shape[0], background_shape[1], 3),
dtype='uint8')
padding_object_area[row:row + object_area_shape[0],
col:col + object_area_shape[1],
0] = object_area[:, :, 0]
padding_object_area[row:row + object_area_shape[0],
col:col + object_area_shape[1],
1] = object_area[:, :, 1]
padding_object_area[row:row + object_area_shape[0],
col:col + object_area_shape[1],
2] = object_area[:, :, 2]
difference = sum(sum(sum(object_area))) - sum(
sum(sum(padding_object_area)))
print(difference)
t_mask = np.where(padding_mask == 1, 0, 1)
background[:, :, 0] = background[:, :, 0] * t_mask
background[:, :, 1] = background[:, :, 1] * t_mask
background[:, :, 2] = background[:, :, 2] * t_mask
# plt.figure('选出background上的区域')
# plt.imshow(background)
# plt.show()
background = background + padding_object_area
# plt.figure('贴好的区域')
# plt.imshow(padding_object_area)
# plt.show()
background = Image.fromarray(background, 'RGB')
tamper_image.append(background)
tamper_mask.append(padding_mask)
background = orignal_bg
except Exception as e:
traceback.print_exc()
else:
# 暂时先空着
pass
for index, item in enumerate(tamper_image):
difference_8 = mask_to_outeedge(tamper_mask[index])
difference_8_dilation = dilation.binary_dilation(
difference_8, np.ones((3, 3)))
difference_8_dilation = np.where(difference_8_dilation == True, 1, 0)
double_edge_candidate = difference_8_dilation + tamper_mask[index]
double_edge = np.where(double_edge_candidate == 2, 1, 0)
ground_truth = np.where(double_edge == 1, 255, 0) + np.where(
difference_8 == 1, 100, 0) + np.where(tamper_mask[index] == 1, 50,
0) # 所以内侧边缘就是100的灰度值
tamper_gt.append(ground_truth)
try:
mask = Image.fromarray(tamper_mask[index])
except Exception as e:
print('mask to Image error', e)
# 在这里的时候，mask foreground background 尺寸都是一致的了，poisson融合时，offset置为0
foreground = Image.fromarray(padding_object_area)
background = Image.fromarray(orignal_bg)
mask = padding_mask
try:
poisson_foreground = cv2.cvtColor(
np.asarray(foreground.convert('RGB')), cv2.COLOR_RGB2BGR)
poisson_background = cv2.cvtColor(np.asarray(background),
cv2.COLOR_RGB2BGR)
poisson_mask = np.asarray(mask)
poisson_mask = np.where(poisson_mask == 1, 255, 0)
poisson_fusion_image = poisson_image_editing.poisson_fusion(
poisson_foreground, poisson_background, poisson_mask)
poisson_fusion_image = Image.fromarray(
cv2.cvtColor(poisson_fusion_image, cv2.COLOR_BGR2RGB))
tamper_poisson.append(poisson_fusion_image)
return tamper_image, tamper_poisson, tamper_gt
except Exception as e:
traceback.print_exc()
