def LoadImages(self):
"""处理每一个训练的image.
处理每一个image，来初始化初始化self.images(存入所有训练图像矩阵)，用mean_shape缩放得到的
landmark初始化self.initLandmarks,用self.origLandmarks(通过读取.pts文件获得)来初始化self.gtLandmarks（gr-
ound truth landmark）
"""
pass
print('Loading Images......')
self.imgs = []
self.initLandmarks = []
self.gtLandmarks = []
for i in range(len(self.filenames)):
print('Loading ......%dth image.....'%(i))
# ndimage.imread类似matlab的imread函数，对于RGB图像，返回height*width*3的矩阵， 对于灰度图返回height*width
img = ndimage.imread(self.filenames[i])
# 如果self.color为true保持rgb图不变灰度图叠加三次，结果为height*width*3 的img
# 如果self.color为false保持灰度图不变rgb图rgb取平均，结果为height*width的img
if self.color:
if len(img.shape) == 2:
# 类似于np.concarenate
img = np.dstack((img, img, img))
else:
if len(img.shape) > 2:
# 转换为灰度图像，将RGB三值取平均
img = np.mean(img, axis=2)
img = img.astype(np.uint8)
if self.mirrors[i]:
# 对需要做特征变换的图像，特征点做一个镜像变换
self.origLandmarks[i] = utils.mirrorShape(self.origLandmarks[i], img.shape)
# 对图像做镜像变换，fliplr即数组左右翻转，
img = np.fliplr(img)
# 如果self.color为true，则调整矩阵维度先后，变为3*height*width
# 如果self.color为false,则增加一个维度， 变为1*height*width
if self.color:
img = np.transpose(img, (2, 0, 1))
else:
img = img[np.newaxis]
groundTruth = self.origLandmarks[i]
if self.initialization == 'rect':
bestFit = utils.bestFitRect(groundTruth, self.meanShape)
elif self.initialization == 'similarity':
bestFit = utils.bestFit(groundTruth, self.meanShape)
elif self.initialization == 'box':
bestFit = utils.bestFitRect(groundTruth, self.meanShape, box=self.boundingBoxes[i])
self.imgs.append(img)
self.initLandmarks.append(bestFit)
self.gtLandmarks.append(groundTruth)
self.initLandmarks = np.array(self.initLandmarks)
self.gtLandmarks = np.array(self.gtLandmarks)
print('Loading Images finished')
def landmark_to_gtlandmark(landmark, meanshape, box=None, mode='rect'):
best_fit = []
if mode == 'rect':
best_fit = bestFitRect(landmark, meanshape)
elif mode == 'similarity':
best_fit = bestFit(landmark, meanshape)
elif mode == 'box':
best_fit = bestFitRect(landmark, meanshape, box)
return best_fit
def LoadImages(self):
self.imgs = []
self.initLandmarks = []
self.gtLandmarks = []
for i in range(len(self.filenames)):
# 이 코드는 잘 작성되지 않았으며 self.color는 실제적인 의미가 없습니다. <<< 这段代码写得不好，self.color并没有实际意义
img = ndimage.imread(self.filenames[i])
if self.color:
if len(img.shape) == 2:
img = np.dstack((img, img, img))
else:
# img = ndimage.imread(self.filenames[i], mode='L')
if len(img.shape) > 2:
img = np.mean(img, axis=2)
img = img.astype(np.uint8)
if self.mirrors[i]:
self.origLandmarks[i] = utils.mirrorShape(
self.origLandmarks[i], img.shape)
img = np.fliplr(img)
if not self.color:
# img = np.transpose(img, (2, 0, 1))
# else:
img = img[
np.
newaxis] # Img는 모양 (H, W)에서 모양 (1, H, W)로 바뀝니다. <<< img从shape(H,W)变成shape(1,H,W)
# img = np.transpose(img, (1, 2, 0))
groundTruth = self.origLandmarks[i]
if self.initialization == 'rect':
bestFit = utils.bestFitRect(
groundTruth, self.meanShape
) # 랜드 마크에 의해 결정된 상자에 도구 모양을 맞추기 만하면됩니다. <<< 仅仅把meanshape适应进入由landmark确定的框中
elif self.initialization == 'similarity':
bestFit = utils.bestFit(
groundTruth, self.meanShape
) # gt에 대한 meanShape의 최상의 변환을 찾아 변환하십시오. <<< 找到meanShape到gt的最优变换，并变换之
elif self.initialization == 'box':
bestFit = utils.bestFitRect(
groundTruth, self.meanShape, box=self.boundingBoxes[i]
) # 감지 된 bbx에 의해 결정된 상자에 도구 모양을 맞 춥니 다. <<< 仅仅把meanshape适应进入由检测到的bbx确定的框中
self.imgs.append(img)
self.initLandmarks.append(bestFit)
self.gtLandmarks.append(groundTruth)
self.initLandmarks = np.array(self.initLandmarks)
self.gtLandmarks = np.array(self.gtLandmarks)
def LoadImages(self):
self.imgs = []
self.initLandmarks = []
self.gtLandmarks = []
for i in range(len(self.filenames)): # 读取文件夹下面的每个图片
img = ndimage.imread(self.filenames[i]) # 图片shape：539 * 410 * 3
if self.color:
if len(img.shape) == 2:
img = np.dstack((img, img, img))
else:
if len(img.shape) > 2: # img的信息坐标轴个数，row，column，channel共3个坐标轴
img = np.mean(
img, axis=2
) # Compute the arithmetic mean along the specified axis.将rgb数据求平均，融合成新的数据
img = img.astype(np.uint8)
if self.mirrors[i]:
self.origLandmarks[i] = utils.mirrorShape(
self.origLandmarks[i], img.shape)
img = np.fliplr(img)
if self.color:
img = np.transpose(img, (2, 0, 1))
else:
img = img[
np.
newaxis] # 在第一维增加一个新的坐标轴，shape由539 * 410，变为1 * 539 * 410
groundTruth = self.origLandmarks[i] # 取到
if self.initialization == 'rect':
bestFit = utils.bestFitRect(
groundTruth, self.meanShape) # 返回由平均特征点经过缩放和偏移，生成的初始特征点坐标。
elif self.initialization == 'similarity':
bestFit = utils.bestFit(groundTruth, self.meanShape)
elif self.initialization == 'box':
bestFit = utils.bestFitRect(groundTruth,
self.meanShape,
box=self.boundingBoxes[i])
self.imgs.append(img) # 把文件夹中的图片依次加入list中
self.initLandmarks.append(
bestFit) # 把文件夹中的图片对应的初始landmark坐标依次加入list中
self.gtLandmarks.append(groundTruth)
if i % 500 == 0:
print("the %dth pic in func LoadImages()" % i)
self.initLandmarks = np.array(self.initLandmarks)
self.gtLandmarks = np.array(self.gtLandmarks)
def LoadImages(self):
self.imgs = []
self.initLandmarks = []
self.gtLandmarks = []
for i in range(len(self.filenames)):
# 这段代码写得不好，self.color并没有实际意义
img = ndimage.imread(self.filenames[i])
if self.color:
if len(img.shape) == 2:
img = np.dstack((img, img, img))
else:
# img = ndimage.imread(self.filenames[i], mode='L')
if len(img.shape) > 2:
img = np.mean(img, axis=2)
img = img.astype(np.uint8)
if self.mirrors[i]:
self.origLandmarks[i] = utils.mirrorShape(
self.origLandmarks[i], img.shape)
img = np.fliplr(img)
if not self.color:
# img = np.transpose(img, (2, 0, 1))
# else:
img = img[np.newaxis] # img从shape(H,W)变成shape(1,H,W)
# img = np.transpose(img, (1, 2, 0))
groundTruth = self.origLandmarks[i]
if self.initialization == 'rect':
# 仅仅把meanshape适应进入由landmark确定的框中
bestFit = utils.bestFitRect(groundTruth, self.meanShape)
elif self.initialization == 'similarity':
# 找到meanShape到gt的最优变换，并变换之
bestFit = utils.bestFit(groundTruth, self.meanShape)
elif self.initialization == 'box':
# 仅仅把meanshape适应进入由检测到的bbx确定的框中
bestFit = utils.bestFitRect(groundTruth,
self.meanShape,
box=self.boundingBoxes[i])
self.imgs.append(img)
self.initLandmarks.append(bestFit)
self.gtLandmarks.append(groundTruth)
self.initLandmarks = np.array(self.initLandmarks)
self.gtLandmarks = np.array(self.gtLandmarks)
def LoadImages(self):
self.imgs = []
self.initLandmarks = []
self.gtLandmarks = []
for i in range(len(self.filenames)):
img = ndimage.imread(self.filenames[i])
if self.color:
if len(img.shape) == 2:
img = np.dstack((img, img, img))
else:
if len(img.shape) > 2:
img = np.mean(img, axis=2)
img = img.astype(np.uint8)
if self.mirrors[i]:
self.origLandmarks[i] = utils.mirrorShape(
self.origLandmarks[i], img.shape)
img = np.fliplr(img)
if self.color:
img = np.transpose(img, (2, 0, 1))
else:
img = img[np.newaxis]
groundTruth = self.origLandmarks[i]
if self.initialization == 'rect':
bestFit = utils.bestFitRect(groundTruth, self.meanShape)
elif self.initialization == 'similarity':
bestFit = utils.bestFit(groundTruth, self.meanShape)
elif self.initialization == 'box':
bestFit = utils.bestFitRect(groundTruth,
self.meanShape,
box=self.boundingBoxes[i])
self.imgs.append(img)
self.initLandmarks.append(bestFit)
self.gtLandmarks.append(groundTruth)
self.initLandmarks = np.array(self.initLandmarks)
self.gtLandmarks = np.array(self.gtLandmarks)
def get_landmarks(bb, img, model):
landmarks = None
print("bb: ", bb)
initLandmarks = utils.bestFitRect(None, model.initLandmarks,
[bb[0], bb[1], bb[2], bb[3]])
landmarks, confidence = model.processImg(img[np.newaxis], initLandmarks)
print("conf: ", confidence)
if confidence < 0.3:
landmarks = None
return landmarks
def alignment(self, bb, img):
landmarks = None
initLandmarks = utils.bestFitRect(None, self.model.initLandmarks, [bb[0], bb[1], bb[2], bb[3]])
landmarks = self.model.processImg(img[np.newaxis], initLandmarks)
"""
landmarks, confidence = self.model.processImg(img[np.newaxis], initLandmarks)
print("alignment conf: ", confidence)
if confidence < 0.9:
landmarks = None
"""
return landmarks
def faceKeypointsDetector(faceImage, model):
img = faceImage.astype(np.uint8)
minX = 0
maxX = faceImage.shape[1]
minY = 0
maxY = faceImage.shape[0]
initLandmarks = utils.bestFitRect(None, model.initLandmarks,
[minX, minY, maxX, maxY])
if model.confidenceLayer:
landmarks, confidence = model.processImg(img[np.newaxis],
initLandmarks)
if confidence < 0.1:
reset = True
else:
landmarks = model.processImg(img[np.newaxis], initLandmarks)
landmarks = landmarks.astype(np.int32)
for i in range(landmarks.shape[0]):
cv2.circle(faceImage, (landmarks[i, 0], landmarks[i, 1]), 2,
(0, 255, 0))
cv2.imshow("faceKeypoints", faceImage)
cv2.waitKey(1)
# reset = True
landmarks = None
# if reset:
rects = cascade.detectMultiScale(gray_img, scaleFactor=1.2, minNeighbors=3, minSize=(50, 50))
for rect in rects:
tl_x = rect[0]
tl_y = rect[1]
br_x = tl_x + rect[2]
br_y = tl_y + rect[3]
cv2.rectangle(color_img, (tl_x, tl_y), (br_x, br_y), (255, 0, 0))
initLandmarks = utils.bestFitRect(None, model.initLandmarks, [tl_x, tl_y, br_x, br_y])
if model.confidenceLayer:
landmarks, confidence = model.processImg(gray_img[np.newaxis], initLandmarks)
if confidence < 0.1:
reset = True
else:
landmarks = model.processImg(gray_img[np.newaxis], initLandmarks)
landmarks = landmarks.astype(np.int32)
for i in range(landmarks.shape[0]):
cv2.circle(color_img, (landmarks[i, 0], landmarks[i, 1]), 2, (0, 255, 0))
cv2.imshow("image", color_img)
key = cv2.waitKey(0)
img = np.mean(vis, axis=2).astype(np.uint8)
else:
img = vis.astype(np.uint8)
if reset:
rects = cascade.detectMultiScale(img,
scaleFactor=1.2,
minNeighbors=3,
minSize=(50, 50))
if len(rects) > 0:
minX = rects[0][0]
maxX = rects[0][0] + rects[0][2]
minY = rects[0][1]
maxY = rects[0][1] + rects[0][3]
cv2.rectangle(vis, (minX, minY), (maxX, maxY), (255, 0, 0))
initLandmarks = utils.bestFitRect(None, model.initLandmarks,
[minX, minY, maxX, maxY])
reset = False
if model.confidenceLayer:
landmarks, confidence = model.processImg(
img[np.newaxis], initLandmarks)
if confidence < 0.1:
reset = True
else:
landmarks = model.processImg(img[np.newaxis], initLandmarks)
landmarks = landmarks.astype(np.int32)
for i in range(landmarks.shape[0]):
cv2.circle(vis, (landmarks[i, 0], landmarks[i, 1]), 2,
(0, 255, 0))
else:
initLandmarks = utils.bestFitRect(landmarks, model.initLandmarks)
img = cv2.imread(filenames[i])
imgColor = np.copy(img[:, :, [2, 1, 0]])
if len(img.shape) > 2:
img = np.mean(img, axis=2)
faceHeight = img.shape[0] * imageHeightFraction
faceWidth = faceHeight
center = np.array(img.shape) / 2
box = [
center[1] - faceWidth / 2, center[0] - faceHeight / 2,
center[1] + faceWidth / 2, center[0] + faceHeight / 2
]
#first step
initLandmarks = utils.bestFitRect([], network.initLandmarks, box)
firstStepLandmarks = network.processImg(img[np.newaxis], initLandmarks)
#second step
normImg, transform = network.CropResizeRotate(img[np.newaxis],
firstStepLandmarks)
normFirstStepLandmarks = np.dot(firstStepLandmarks,
transform[0]) + transform[1]
initLandmarks2 = utils.bestFitRect(normFirstStepLandmarks,
network.initLandmarks)
finalLandmarks = network.processImg(normImg, initLandmarks2)
finalLandmarks = np.dot(finalLandmarks - transform[1],
np.linalg.inv(transform[0]))
baseName = ntpath.basename(filenames[i])[:-4]
