#!/usr/bin/python2
# -*- coding: utf-8 -*-
# vim: expandtab:ts=4:sw=4
# --------------------------------------------------------
import cv2
import copy
import time
import numpy as np
from numpy import random
import colorsys
def bbox_area(bbox):
return (bbox[:, 2] - bbox[:, 0]) * (bbox[:, 3] - bbox[:, 1])
def clip_box(bbox, labels, clip_box, alpha):
"""Clip the bounding boxes to the borders of an image
"""
ar_ = (bbox_area(bbox))
delta_area = ((ar_ - bbox_area(clip_box)) / ar_)
mask = (delta_area < (1 - alpha)).astype(int)
bbox = bbox[mask == 1, :]
new_labels = labels[mask == 1]
return bbox, new_labels
class ConvertFromInts(object):
def __call__(self, img, boxes=None, labels=None):
return img.astype(np.float32), boxes.astype(np.float32), labels
class ConvertFromFloats(object):
def __call__(self, img, boxes=None, labels=None):
return img.astype(np.int32), boxes.astype(np.int32), labels
class Normalization(object):
"""Normalization: Pixel Value (0, 255) to (0, 1)"""
def __call__(self, img, boxes=None, labels=None):
img /= 255
return img, boxes, labels
class InverseNormalization(object):
"""Inverse Normalization: Pixel Value (0, 1) to (0, 255)"""
def __call__(self, img, boxes=None, labels=None):
img *= 255
img = np.clip(img, 0, 255)
return img, boxes, labels
class SubtractMeans(object):
def __init__(self, mean):
self.mean = mean
def __call__(self, img, boxes=None, labels=None):
img = img.astype(np.float32)
img -= self.mean
return img.astype(np.float32), boxes, labels
class ToAbsoluteCoords(object):
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
boxes[:, 0] *= width
boxes[:, 2] *= width
boxes[:, 1] *= height
boxes[:, 3] *= height
return img, boxes, labels
class ToPercentCoords(object):
def __call__(self, img, boxes=None, labels=None):
height, width, labels = img.shape
boxes[:, 0] /= width
boxes[:, 2] /= width
boxes[:, 1] /= height
boxes[:, 3] /= height
return img, boxes, labels
class HorizontalFlip(object):
def __call__(self, img, boxes=None, labels=None):
img = cv2.flip(img, 1)
height, width, channels = img.shape
boxes[:, 0] = width - boxes[:, 0]
boxes[:, 2] = width - boxes[:, 2]
return img, boxes, labels
class RandomHorizontalFlip(object):
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
img = cv2.flip(img, 1)
height, width, channels = img.shape
boxes[:, 0] = width - boxes[:, 0]
boxes[:, 2] = width - boxes[:, 2]
return img, boxes, labels
class VehicleFlip(object):
def __call__(self, img, boxes=None, labels=None):
img = cv2.flip(img, 0)
height, width, channels = img.shape
boxes[:, 1] = height - boxes[:, 1]
boxes[:, 3] = height - boxes[:, 3]
return img, boxes, labels
class RandomVehicleFlip(object):
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
img = cv2.flip(img, 0)
height, width, channels = img.shape
boxes[:, 1] = height - boxes[:, 1]
boxes[:, 3] = height - boxes[:, 3]
return img, boxes, labels
class Resize(object):
"""Opencv resize"""
def __init__(self, width=300, height=300):
self.width = width
self.height = height
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
scale_x = float(self.width) / float(width)
scale_y = float(self.height) / float(height)
img= cv2.resize(img, None, fx = scale_x, fy = scale_y, interpolation=cv2.INTER_CUBIC)
boxes[:, 0] *= scale_x
boxes[:, 2] *= scale_x
boxes[:, 1] *= scale_y
boxes[:, 3] *= scale_y
return img, boxes, labels
class LetterBox(object):
"""Resize the image in accordance to 'image_letter_box' function in darknet
The aspect ratio is maintained."""
def __init__(self, width=600, height=600):
self.width = width
self.height = height
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
scale_x = float(self.width) / float(width)
scale_y = float(self.height) / float(height)
if scale_x > scale_y:
scale = scale_y
else:
scale = scale_x
new_w = int(width * scale)
new_h = int(height * scale)
resized_img = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_CUBIC)
new_img = np.full((self.height, self.width, 3), 128)
new_img[(self.height - new_h)//2:(self.height - new_h)//2 + new_h,(self.width - new_w)//2:(self.width - new_w)//2 + new_w, :] = resized_img
boxes *= scale
boxes[:, 0] += (self.width - new_w)//2
boxes[:, 2] += (self.width - new_w)//2
boxes[:, 1] += (self.height - new_h)//2
boxes[:, 3] += (self.height - new_h)//2
return new_img, boxes, labels
class ConvertRGB2HSV(object):
def __call__(self, img, boxes=None, labels=None):
img = cv2.cvtColor(img, cv2.COLOR_RGB2HSV)
return img, boxes, labels
class ConvertHSV2RGB(object):
def __call__(self, img, boxes, labels):
img = cv2.cvtColor(img, cv2.COLOR_HSV2RGB)
return img, boxes, labels
class RandomSaturation(object):
"""Transfrom the image in HSV color space"""
def __init__(self, saturation=0.5):
assert saturation > 0.0 and saturation < 1.0
self.saturation = saturation
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
img[:, :, 1] *= random.uniform(-self.saturation, self.saturation)
return img, boxes, labels
class RandomHue(object):
"""Transfrom the image in HSV color space"""
def __init__(self, hue=18.0):
assert hue >= 0.0 and hue <= 360.0
self.hue = hue
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
img[:, :, 0] += random.uniform(-self.hue, self.hue)
img[:, :, 0] = np.clip(img[:, :, 0], 0, 360)
return img, boxes, labels
class RandomBrightness(object):
"""Tranfrom the image in RGB color space"""
def __init__(self, brightness=32):
assert brightness > 0.0 and brightness < 255.0
self.brightness = brightness
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
img += random.uniform(-self.brightness, self.brightness)
img = np.clip(img, 0.0, 255.0)
return img, boxes, labels
class RandomContrast(object):
"""Tranfrom the image in RGB color space"""
def __init__(self, contrast=0.9):
assert contrast > 0.0 and contrast < 1.0
self.contrast = contrast
def __call__(self, img, boxes, labels):
if random.randint(2):
img *= random.uniform(-self.contrast, self.contrast)
img = np.clip(img, 0.0, 255.0)
return img, boxes, labels
class PhotometricDistort(object):
def __init__(self):
self.pd = Compose([
RandomContrast(),
ConvertRGB2HSV(),
RandomSaturation(),
RandomHue(),
ConvertHSV2RGB(),
RandomBrightness()
])
def __call__(self, img, boxes=None, labels=None):
return self.pd(img, boxes, labels)
class AddGaussNoise(object):
def __init__(self, sigma=2):
self.sigma = sigma
def __call__(self, img, boxes=None, labels=None):
sigma = random.uniform(0, self.sigma)
img = cv2.GaussianBlur(img, (5,5), sigmaX=sigma)
return img, boxes, labels
class Expand(object):
"""Expand the image
mean: the pixel value of the expand area.
ratio: the max ratio of the orgin image width/height with expanded image width.height """
def __init__(self, mean=128, ratio=1.2):
self.ratio = ratio
self.mean = mean
def __call__(self, img, boxes=None, labels=None):
if random.randint(2):
return img, boxes, labels
ratio = random.uniform(1, self.ratio)
height, width, channels = img.shape
new_w = int(width * ratio)
new_h = int(height * ratio)
if new_w == width or new_h == height:
return img, boxes, labels
left = random.randint(0, new_w - width)
top = random.randint(0, new_h - height)
expand_image = np.full((new_h, new_w, 3), self.mean)
expand_image[top:top+height, left:left+width, :] = img
img = expand_image
boxes[:, 0] += left
boxes[:, 2] += left
boxes[:, 1] += top
boxes[:, 3] += top
return img, boxes, labels
class Four_Point_Crop(object):
"""crop_x and crop_y range from -0.5 to 0.5
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
crop_x > 0 , crop_y > 0 crop image upper left part
crop_x > 0 , crop_y < 0 crop image Lower left part
crop_x < 0 , crop_y > 0 crop image upper right part
crop_x < 0 , crop_y < 0 crop image right lower part
"""
def __init__(self, crop_x=0.2, crop_y=0.2):
assert crop_x < 0.5 and crop_x > -0.5
assert crop_y < 0.5 and crop_y > -0.5
self.crop_x = crop_x
self.crop_y = crop_y
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
new_w = int(width * (1 - abs(self.crop_x)))
new_h = int(height * (1 - abs(self.crop_y)))
[left, top, right, bottom] = [0, 0, width, height]
if self.crop_x >= 0 and self.crop_y >= 0:
right = new_w
bottom = new_h
if self.crop_x <= 0 and self.crop_y <= 0:
left = width - new_w
top = height - new_h
if self.crop_x >=0 and self.crop_y <= 0:
top = height -new_h
right = new_w
if self.crop_x <= 0 and self.crop_y >= 0:
left = width - new_w
bottom = new_h
new_img = img[top:bottom, left:right, :]
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], left)
new_boxes[:, 1] = np.maximum(boxes[:, 1], top)
new_boxes[:, 2] = np.minimum(boxes[:, 2], right)
new_boxes[:, 3] = np.minimum(boxes[:, 3], bottom)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
boxes[:, 0] -= left
boxes[:, 1] -= top
boxes[:, 2] -= left
boxes[:, 3] -= top
return new_img, boxes, labels
class CenterCrop(object):
"""crop_x and crop_y range from 0 to 0.5
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, crop_x=0.5, crop_y=0.5):
assert crop_x < 0.5 and crop_x > 0
assert crop_y < 0.5 and crop_y > 0
self.crop_x = crop_x
self.crop_y = crop_y
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
new_w = int(width * (1 - abs(self.crop_x)))
new_h = int(height * (1 - abs(self.crop_y)))
left = (width - new_w) // 2
top = (height -new_h) // 2
right = left + new_w
bottom = top + new_h
new_img = img[top:bottom, left:right, :]
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], left)
new_boxes[:, 1] = np.maximum(boxes[:, 1], top)
new_boxes[:, 2] = np.minimum(boxes[:, 2], right)
new_boxes[:, 3] = np.minimum(boxes[:, 3], bottom)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
boxes[:, 0] -= left
boxes[:, 1] -= top
boxes[:, 2] -= left
boxes[:, 3] -= top
return new_img, boxes, labels
class RandomFour_Point_Crop(object):
"""crop_x and crop_y range from 0 to 0.5
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, crop_x=0.2, crop_y=0.2):
assert crop_x < 0.5 and crop_x > 0
assert crop_y < 0.5 and crop_y > 0
self.crop_x = crop_x
self.crop_y = crop_y
def __call__(self, img, boxes=None, labels=None):
crop_x = random.uniform(-self.crop_x, self.crop_x)
crop_y = random.uniform(-self.crop_y, self.crop_y)
height, width, channels = img.shape
new_w = int(width * (1 - abs(crop_x)))
new_h = int(height * (1 - abs(crop_y)))
[left, top, right, bottom] = [0, 0, width, height]
if crop_x >= 0 and crop_y >= 0:
right = new_w
bottom = new_h
if crop_x <= 0 and crop_y <= 0:
left = width - new_w
top = height - new_h
if crop_x >=0 and crop_y <= 0:
top = height -new_h
right = new_w
if crop_x <= 0 and crop_y >= 0:
left = width - new_w
bottom = new_h
new_img = img[top:bottom, left:right, :]
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], left)
new_boxes[:, 1] = np.maximum(boxes[:, 1], top)
new_boxes[:, 2] = np.minimum(boxes[:, 2], right)
new_boxes[:, 3] = np.minimum(boxes[:, 3], bottom)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
boxes[:, 0] -= left
boxes[:, 1] -= top
boxes[:, 2] -= left
boxes[:, 3] -= top
return new_img, boxes, labels
class RandomCenterCrop(object):
"""crop_x and crop_y range from 0 to 0.5
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, crop_x=0.5, crop_y=0.5):
assert crop_x < 0.5 and crop_x > 0
assert crop_y < 0.5 and crop_y > 0
self.crop_x = crop_x
self.crop_y = crop_y
def __call__(self, img, boxes=None, labels=None):
crop_x = random.uniform(-self.crop_x, self.crop_x)
crop_y = random.uniform(-self.crop_y, self.crop_y)
height, width, channels = img.shape
new_w = int(width * (1 - abs(crop_x)))
new_h = int(height * (1 - abs(crop_y)))
left = (width - new_w) // 2
top = (height -new_h) // 2
right = left + new_w
bottom = top + new_h
new_img = img[top:bottom, left:right, :]
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], left)
new_boxes[:, 1] = np.maximum(boxes[:, 1], top)
new_boxes[:, 2] = np.minimum(boxes[:, 2], right)
new_boxes[:, 3] = np.minimum(boxes[:, 3], bottom)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
boxes[:, 0] -= left
boxes[:, 1] -= top
boxes[:, 2] -= left
boxes[:, 3] -= top
return new_img, boxes, labels
class RandomCrop(object):
"""crop_x and crop_y range from 0 to 0.5
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, crop_x=0.4, crop_y=0.4):
self.crop_x = crop_x
self.crop_y = crop_y
def __call__(self, img, boxes=None, labels=None):
if random.randint(0, 5) == 0:
randomcrop = Compose([RandomCenterCrop(self.crop_x, self.crop_y)])
else:
randomcrop = Compose([RandomFour_Point_Crop(self.crop_x, self.crop_y)])
img, boxes, labels = randomcrop(img, boxes, labels)
return img, boxes, labels
class Rotate(object):
"""Rotates an image.
Note: the big angle make the lower-left corner of the image lost.
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, angle=3):
assert angle > -360 and angle < 360
self.angle = angle
def __call__(self, img, boxes=None, labels=None):
height, width, channels = img.shape
center_x, center_y = width // 2, height // 2
M = cv2.getRotationMatrix2D((center_x, center_y), self.angle, 1.0)
cos = np.abs(M[0, 0])
sin = np.abs(M[0, 1])
nW = int(height * sin + width * cos)
nH = int(height * cos + width * sin)
M[0, 2] += (nW / 2) - center_x
M[1, 2] += (nH / 2) - center_y
img = cv2.warpAffine(img, M, (nW, nH))
corners = np.hstack((boxes[:, 0].reshape(-1, 1), boxes[:, 1].reshape(-1, 1),
boxes[:, 2].reshape(-1, 1), boxes[:, 1].reshape(-1, 1),
boxes[:, 0].reshape(-1, 1), boxes[:, 3].reshape(-1, 1),
boxes[:, 2].reshape(-1, 1), boxes[:, 3].reshape(-1, 1)))
corners = corners.reshape(-1, 2)
corners = np.hstack((corners, np.ones((corners.shape[0], 1), dtype=type(corners[0][0]))))
corners = np.dot(M, corners.T).T
corners = corners.reshape(-1, 8)
corners_x = corners[:, [0, 2, 4, 6]]
corners_y = corners[:, [1, 3, 5, 7]]
xmin = np.min(corners_x, 1).reshape(-1, 1)
ymin = np.min(corners_y, 1).reshape(-1, 1)
xmax = np.max(corners_x, 1).reshape(-1, 1)
ymax = np.max(corners_y, 1).reshape(-1, 1)
boxes = np.hstack((xmin, ymin, xmax, ymax))
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], 0)
new_boxes[:, 1] = np.maximum(boxes[:, 1], 0)
new_boxes[:, 2] = np.minimum(boxes[:, 2], width)
new_boxes[:, 3] = np.minimum(boxes[:, 3], height)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
return img, boxes, labels
class RandomRotate(object):
"""Rotates an image.
Note: the big angle make the lower-left corner of the image lost.
Bounding boxes which have an area of less than 25% in the remaining in the
transformed image is dropped.
"""
def __init__(self, angle=10):
assert angle > 0 and angle < 360
self.angle = angle
def __call__(self, img, boxes=None, labels=None):
if len(boxes) == 0 :
return img, boxes, labels
if random.randint(2):
return img, boxes, labels
angle = random.uniform(-self.angle, self.angle)
print('randomangle', angle)
height, width, channels = img.shape
center_x, center_y = width // 2, height // 2
M = cv2.getRotationMatrix2D((center_x, center_y), angle, 1.0)
cos = np.abs(M[0, 0])
sin = np.abs(M[0, 1])
nW = int(height * sin + width * cos)
nH = int(height * cos + width * sin)
M[0, 2] += (nW / 2) - center_x
M[1, 2] += (nH / 2) - center_y
img = img.astype(np.uint8)
img = cv2.warpAffine(img, M, (nW, nH))
corners = np.hstack((boxes[:, 0].reshape(-1, 1), boxes[:, 1].reshape(-1, 1),
boxes[:, 2].reshape(-1, 1), boxes[:, 1].reshape(-1, 1),
boxes[:, 0].reshape(-1, 1), boxes[:, 3].reshape(-1, 1),
boxes[:, 2].reshape(-1, 1), boxes[:, 3].reshape(-1, 1)))
corners = corners.reshape(-1, 2)
corners = np.hstack((corners, np.ones((corners.shape[0], 1), dtype=type(corners[0][0]))))
corners = np.dot(M, corners.T).T
corners = corners.reshape(-1, 8)
corners_x = corners[:, [0, 2, 4, 6]]
corners_y = corners[:, [1, 3, 5, 7]]
xmin = np.min(corners_x, 1).reshape(-1, 1)
ymin = np.min(corners_y, 1).reshape(-1, 1)
xmax = np.max(corners_x, 1).reshape(-1, 1)
ymax = np.max(corners_y, 1).reshape(-1, 1)
boxes = np.hstack((xmin, ymin, xmax, ymax))
new_boxes = copy.deepcopy(boxes)
new_boxes[:, 0] = np.maximum(boxes[:, 0], 0)
new_boxes[:, 1] = np.maximum(boxes[:, 1], 0)
new_boxes[:, 2] = np.minimum(boxes[:, 2], width)
new_boxes[:, 3] = np.minimum(boxes[:, 3], height)
boxes, labels = clip_box(boxes, labels, new_boxes, 0.25)
return img, boxes, labels
class Compose(object):
def __init__(self, transfroms):
self.transfroms = transfroms
def __call__(self, img, boxes=None, labels=None):
for t in self.transfroms:
img, boxes, labels = t(img, boxes, labels)
print(str(t), boxes)
return img, boxes, labels
class Augmentation(object):
def __init__(self, size=300, mean=(104, 117, 123)):
self.mean = mean
self.size = size
self.augment = Compose([
# ConvertFromInts(),
# Resize(),
# PhotometricDistort(),
AddGaussNoise(),
# Expand(),
# RandomCrop(),
# RandomRotate(),
# RandomHorizontalFlip(),
# ConvertFromFloats(),
HorizontalFlip(),
VehicleFlip()
])
def __call__(self, img, boxes, labels):
return self.augment(img, boxes, labels)
def create_unique_color_float(tag, hue_step=0.41):
h, v = (tag * hue_step) % 1, 1. - (int(tag * hue_step) % 4) / 5.
r, g, b = colorsys.hsv_to_rgb(h, 1., v)
return r, g, b
def create_unique_color_uchar(tag, hue_step=0.41):
r, g, b = create_unique_color_float(tag, hue_step)
return int(255 * r), int(255 * g), int(255 * b)
def rectangle(x, y, w, h, alpha, thickness, color, copyedImage, label=None):
pt1 = int(x), int(y)
pt2 = int(x + w), int(y + h)
cv2.rectangle(copyedImage, pt1, pt2, color, thickness)
# FONT_HERSHEY_TRIPLEX
if label is not None:
text_size = cv2.getTextSize(label, cv2.FONT_HERSHEY_COMPLEX, 0.5, 1)
center = pt1[0] + 5, pt1[1] + 11 + text_size[0][1]
pt2 = pt1[0] + 38 + text_size[0][0], pt1[1] + 14 + \
text_size[0][1]
# cv2.rectangle(copyedImage, pt1, pt2, color, -1)
cv2.rectangle(copyedImage, pt1, pt2, color, 1)
cv2.putText(copyedImage, label, center, cv2.FONT_HERSHEY_COMPLEX, 0.8, (0, 255, 255), 1)
def main():
imgname = '/home/liujunyan/Documents/dep/test/ccc.jpg'
txtname = '/home/liujunyan/Documents/dep/test/ccc.txt'
boxes = []
labels = []
img = cv2.imread(imgname)
(i_h, i_w, _) = img.shape
with open(txtname, 'r') as tf:
lines = tf.readlines()
for line in lines:
lab = line.strip().split(' ')
labels.append(lab[0])
# box : xmin, ymin, xmax, ymax
xmin = (float(lab[1])-float(lab[3])/2.)*i_w
ymin = (float(lab[2])-float(lab[4])/2.)*i_h
xmax = (float(lab[1])+float(lab[3])/2.)*i_w
ymax = (float(lab[2])+float(lab[4])/2.)*i_h
boxes.append([xmin, ymin, xmax, ymax])
boxes = np.array(boxes,dtype=np.float32)
labels = np.array(labels,dtype=np.float32)
aug = Augmentation()
aug(img, boxes, labels)
auged_img, auged_bboxes, auged_labels = aug(img, boxes, labels)
return auged_img, auged_bboxes, auged_labels
if __name__ == "__main__":
main()
Image augmentation图像增广
最后编辑于 ：
©著作权归作者所有,转载或内容合作请联系作者
- 文/潘晓璐 我一进店门，熙熙楼的掌柜王于贵愁眉苦脸地迎上来，“玉大人，你说我怎么就摊上这事。” “怎么了？”我有些...
- 文/花漫 我一把揭开白布。 她就那样静静地躺着，像睡着了一般。 火红的嫁衣衬着肌肤如雪。 梳的纹丝不乱的头发上，一...
- 文/苍兰香墨 我猛地睁开眼，长吁一口气：“原来是场噩梦啊……” “哼！你这毒妇竟也来了？” 一声冷哼从身侧响起，我...
