基於yolo3自定義訓練數據（三）使用imgaug擴大數據集

一、imguag簡介

 

備選參考的圖片擴大框架：kears Imagedatagenerator

參考文檔

https://imgaug.readthedocs.io/en/latest/

https://github.com/aleju/imgaug

python3.7

numpy1.17.0

參數選擇：

• 如果可能，應使用最近鄰插值或線性插值，因為它們比其他選項要快得多。使用插值的大多數增強器提供order參數（0 =最近鄰，1 =線性）或interpolation參數（“最近鄰”，“線性”）。
• keep_size=True在所有更改圖像尺寸的增強器中，默認設置為使用。這很方便，因為它可以確保圖像尺寸不會因擴展而改變。但是，它的確會導致明顯的性能下降，通常不僅僅使帶寬減半。keep_size=False盡可能嘗試 。您仍然可以在擴充后或使用來手動調整圖像的大小KeepSizeByResize(Sequential(<augmenters>))。
• 當增強器提供以用戶定義的方式填充新創建的像素的模式（例如pad_mode=constant，Pad以指定的恆定顏色填充所有填充的像素）時，使用edge代替constant 通常不會帶來明顯的性能損失。

具體的增強器建議：

• 對於存在元素級同級的增強器（例如Multiply和 MultiplyElementwise），元素級增強器通常比非元素級的顯着慢。
• 如果需要模糊處理，AverageBlur是最快的選擇，其次是GaussianBlur。
• 在較粗糙的圖像（例如CoarseDropoutvs Dropout）上運行的增強器可能比其非粗略的兄弟姐妹快得多。
• 對比度歸一化增強器在性能上均具有可比性，但基於直方圖的增強器明顯較慢。
• PiecewiseAffine 是一個非常慢的增幅器，通常應由ElasticTransformation代替，ElasticTransformation可以實現類似的輸出，並且速度要快得多。
• Superpixels是一個相當緩慢的增強器，通常應該包裝起來，例如Sometimes不要經常使用它並降低其性能影響。
• 除天氣FastSnowyLandscape增速器外，其他增速器都相當緩慢，僅在合理時才使用。

圖片

以下數字代表小圖像（64x64x3）和大圖像（224x224x3）。B=1表示的批量大小1，B=128其中一個128。

https://imgaug.readthedocs.io/en/latest/source/performance.html

 

 

二、安裝imguag

官網文檔：https://imgaug.readthedocs.io/en/latest/source/installation.html

在anaconda上安裝

conda config --add channels conda-forge
conda install imgaug

 

 

三、imguag使用方法

https://blog.csdn.net/limiyudianzi/article/details/86497305

https://blog.csdn.net/qq_38451119/article/details/82417412

例一：邊界框編#邊界框

import imgaug as ia import imgaug.augmenters as iaa from imgaug.augmentables.bbs import BoundingBox, BoundingBoxesOnImage ia.seed(1) image = ia.quokka(size=(256, 256)) bbs = BoundingBoxesOnImage([ BoundingBox(x1=65, y1=100, x2=200, y2=150), BoundingBox(x1=150, y1=80, x2=200, y2=130) ], shape=image.shape) seq = iaa.Sequential([ iaa.Multiply((1.2, 1.5)), # change brightness, doesn't affect BBs
 iaa.Affine( translate_px={"x": 40, "y": 60}, scale=(1.2, 1.2) ) # 在x / y軸上平移40 / 60px，縮放到50-70％，影響BB
]) # Augment BBs and images.
image_aug, bbs_aug = seq(image=image, bounding_boxes=bbs) # print coordinates before/after augmentation (see below) # use .x1_int, .y_int, ... to get integer coordinates
for i in range(len(bbs.bounding_boxes)): before = bbs.bounding_boxes[i] after = bbs_aug.bounding_boxes[i] print("BB %d: (%.4f, %.4f, %.4f, %.4f) -> (%.4f, %.4f, %.4f, %.4f)" % ( i, before.x1, before.y1, before.x2, before.y2, after.x1, after.y1, after.x2, after.y2) ) # image with BBs before/after augmentation (shown below)
image_before = bbs.draw_on_image(image, size=2) image_after = bbs_aug.draw_on_image(image_aug, size=2, color=[0, 0, 255]) path_name='F:/esint/smoking_Recognition/test/starzhai'　　#設置自己的保存路徑 
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime())) cv2.imwrite(img_name,image_before) time.sleep(1)　　#以當前時間命名，每隔一秒保存一次圖片 img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime())) cv2.imwrite(img_name,image_after) #能夠進行的操作 平移，放縮，旋轉（？怎么整？）

實現效果：

 

 

例二：熱力圖變換

#熱力圖
import imageio
import numpy as np
import imgaug as ia
import imgaug.augmenters as iaa
from imgaug.augmentables.heatmaps import HeatmapsOnImage
ia.seed(1)
# Load an example image (uint8, 128x128x3).
image = ia.quokka(size=(128, 128), extract="square")
depth = np.linspace(0, 50, 128).astype(np.float32)  # 128 values from 0.0 to 50.0
depth = np.tile(depth.reshape(1, 128), (128, 1))    # change to a horizontal gradient
depth[64-2:64+2, 16:128-16] = 0.75 * 50.0  # line from left to right
depth[16:128-16, 64-2:64+2] = 1.0 * 50.0   # line from top to bottom
depth = HeatmapsOnImage(depth, shape=image.shape, min_value=0.0, max_value=50.0)
depth = depth.avg_pool(2)
# Define our augmentation pipeline.
seq = iaa.Sequential([
    iaa.Dropout([0.05, 0.2]),      # drop 5% or 20% of all pixels
    iaa.Sharpen((0.0, 1.0)),       # sharpen the image
    iaa.Affine(rotate=(-45, 45)),  # rotate by -45 to 45 degrees (affects heatmaps)
    iaa.ElasticTransformation(alpha=50, sigma=5)  # apply water effect (affects heatmaps)
], random_order=True)
# Augment images and heatmaps.
images_aug = []
heatmaps_aug = []
for _ in range(5):
    images_aug_i, heatmaps_aug_i = seq(image=image, heatmaps=depth)
    images_aug.append(images_aug_i)
    heatmaps_aug.append(heatmaps_aug_i)
cells = []
for image_aug, heatmap_aug in zip(images_aug, heatmaps_aug):
    cells.append(image)                                                     # column 1
    cells.append(image_aug)                                                 # column 2
    cells.append(heatmap_aug.draw_on_image(image_aug)[0])                   # column 3
    cells.append(heatmap_aug.draw(size=image_aug.shape[:2])[0])             # column 4
    cells.append(heatmap_aug.draw(size=image_aug.shape[:2], cmap=None)[0])  # column 5
    path_name='F:/esint/smoking_Recognition/test/starzhai'
    img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
    cv2.imwrite(img_name,image)
    time.sleep(1)
    img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
    cv2.imwrite(img_name,image_aug)
    time.sleep(1)
    img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
    cv2.imwrite(img_name,heatmap_aug.draw_on_image(image_aug)[0])
    time.sleep(1)
    img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
    cv2.imwrite(img_name,heatmap_aug.draw(size=image_aug.shape[:2])[0])
    time.sleep(1)
    img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
    cv2.imwrite(img_name,heatmap_aug.draw(size=image_aug.shape[:2], cmap=None)[0])
    time.sleep(1)

voc數據集的擴增方法 

 https://blog.csdn.net/coooo0l/article/details/84492916

 

2.數據變換類別
iaa.AdditiveGaussianNoise(scale=0.2*255)1#加噪聲
iaa.GaussianBlur(sigma=(0.0, 3.0)1#變模糊
iaa.AllChannelsCLAHE(clip_limit=(1, 10))#圖像增強
iaa.Affine(rotate=-45)#左旋轉45度
iaa.Affine(rotate=45)#右旋轉45度
iaa.Affine(scale=(0.4, 0.7))#Scale images to a value of 50 to 150% of their original size變大
iaa.Affine(scale=(1.3, 1.6))#Scale images to a value of 50 to 150% of their original size#縮小
iaa.GammaContrast((0.5, 2.0), per_channel=True)#色相變暗
iaa.Grayscale(alpha=(0.0, 1.0))#色相變灰
iaa.PiecewiseAffine(scale=(0.01, 0.05))#扭曲圖像

 四、批量處理圖片生成對應xml文件詳細實現過程

①默認老鼠圖片以及標注框坐標，輸出10張圖片及其標注框坐標

#邊界框
import imgaug as ia
import imgaug.augmenters as iaa
from imgaug.augmentables.bbs import BoundingBox, BoundingBoxesOnImage
import time
import cv2

#圖像歸一化
ia.seed(1)
image = ia.quokka(size=(256, 256))
bbs = BoundingBoxesOnImage([
    BoundingBox(x1=65, y1=100, x2=200, y2=150),
    BoundingBox(x1=150, y1=80, x2=200, y2=130)
], shape=image.shape)

#aug1=iaa.Affine(rotate=45)#右旋轉45度
#seq = iaa.Sequential([
aug1=iaa.AdditiveGaussianNoise(scale=0.2*255)#加噪聲
aug2=iaa.GaussianBlur(sigma=(0.0, 3.0)) #變模糊
aug3=iaa.AllChannelsCLAHE(clip_limit=(1,10))#圖像增強
aug4=iaa.Affine(rotate=-45)#左旋轉45度
aug5=iaa.Affine(rotate=45)#右旋轉45度
aug6=iaa.Affine(scale=(0.4, 0.7))#Scale images to a value of 50 to 150% of their original size變大
aug7=iaa.Affine(scale=(1.3, 1.6))#Scale images to a value of 50 to 150% of their original size#縮小
aug8=iaa.GammaContrast((2.5, 2.5), per_channel=True)#色相變暗
aug9=iaa.Grayscale(alpha=(0.9))#色相變灰
aug10=iaa.PiecewiseAffine(scale=(0.08))#扭曲圖像
#])

#image_aug, bbs_aug = aug1(image=image, bounding_boxes=bbs)
image1,bbs1=aug1(image=image, bounding_boxes=bbs)
image2,bbs2=aug2(image=image, bounding_boxes=bbs)
image3,bbs3=aug3(image=image, bounding_boxes=bbs)
image4,bbs4=aug4(image=image, bounding_boxes=bbs)
image5,bbs5=aug5(image=image, bounding_boxes=bbs)
image6,bbs6=aug6(image=image, bounding_boxes=bbs)
image7,bbs7=aug7(image=image, bounding_boxes=bbs)
image8,bbs8=aug8(image=image, bounding_boxes=bbs)
image9,bbs9=aug9(image=image, bounding_boxes=bbs)
image10,bbs10=aug10(image=image, bounding_boxes=bbs)

for i in range(len(bbs.bounding_boxes)):
    before = bbs.bounding_boxes[i]
    after = bbs_aug.bounding_boxes[i]
    print("BB %d: (%.4f, %.4f, %.4f, %.4f) -> (%.4f, %.4f, %.4f, %.4f)" % (
        i,
        before.x1, before.y1, before.x2, before.y2,
        after.x1, after.y1, after.x2, after.y2)
    )

#畫標注框
#image_before = bbs.draw_on_image(image, size=2)
#image_after = bbs_aug.draw_on_image(image_aug, size=2, color=[0, 0, 255])
image1 = bbs1.draw_on_image(image1, size=2)
image2 = bbs2.draw_on_image(image2, size=2)
image3 = bbs3.draw_on_image(image3, size=2)
image4 = bbs4.draw_on_image(image4, size=2)
image5 = bbs5.draw_on_image(image5, size=2)
image6 = bbs6.draw_on_image(image6, size=2)
image7 = bbs7.draw_on_image(image7, size=2)
image8 = bbs8.draw_on_image(image8, size=2)
image9 = bbs9.draw_on_image(image9, size=2)
image10 = bbs10.draw_on_image(image10, size=2)

path_name='F:/esint/smoking_Recognition/test/imgaug'

img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image1)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image2)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image3)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image4)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image5)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image6)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image7)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image8)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image9)
time.sleep(1)
img_name = '%s/%s.jpg'%(path_name,time.strftime("%Y%m%d%H%M%S",time.localtime()))
cv2.imwrite(img_name,image10)

 

②更改輸入圖片，更改色階（不要發紫的顏色）

③通過xml文件中的矩形框信息在圖片上畫框

④生成圖片和對應的xml文件

⑤檢測xml文件中的信息是否准確的落在原圖合適的位置上

⑥批量生成圖片及其對應的xml文件（使xml文件名與圖片名一一對應）