【637】一个图片两个标注的图像增强

【529】图像增强（imgaug 包）

　　对于上面关于 imgaug 包的描述，只能针对一套标注图片进行图像增强，不过对于一张图片有两种标注的情况无法实现，需要对两个标注生成一样的变换。　　

　　对于包含楼顶、楼体，需要同时调整变换的时候，内置的方法无法实现，因此我自己建立随机数据来实现，实现思路如下：

对于每一个方法构建随机数，并进行赋值
建立随机数，判断是和增强
对于每一组随机数只产生唯一的增强方案，并同时应用到两个标注上面
这样便产生了一一对应的标注增强方案

　　实现代码：

from imgaug import augmenters as iaa 
from imgaug.augmentables.segmaps import SegmentationMapsOnImage
from PIL import Image 
import numpy as np 
import cv2, os, random 
 
def get_random(min_val, max_val):
    return random.random() * (max_val - min_val) + min_val
 
img_size = 512
 
src_img_dir         = "03_model_dataset_split/01_images/"
src_seg_dir_louding = "03_model_dataset_split/02_labels_louding/"
src_seg_dir_louti   = "03_model_dataset_split/03_labels_louti/"
 
dst_img_dir         = "07_dataset_aug_512_both/01_images/"
dst_seg_dir_louding = "07_dataset_aug_512_both/02_labels_louding/"
dst_seg_dir_louti   = "07_dataset_aug_512_both/03_labels_louti/"
 
for file in os.listdir(src_img_dir):
    src_img_path = os.path.join(src_img_dir, file)
    src_seg_path_louding = os.path.join(src_seg_dir_louding, file)
    src_seg_path_louti = os.path.join(src_seg_dir_louti, file)
     
    img = Image.open(src_img_path)
    img_arr = np.array(img)
     
    seg_louding = Image.open(src_seg_path_louding)
    seg_arr_louding = np.array(seg_louding)
     
    seg_louti = Image.open(src_seg_path_louti)
    seg_arr_louti = np.array(seg_louti)
     
    # segmentation_maps 数据格式需要是 (512, 512, 1)
    # seg_arr 为 (512, 512)，下面是输入 seq 函数的必须格式
    seg_map_louding = SegmentationMapsOnImage(np.expand_dims(seg_arr_louding, axis=-1), shape=(img_size, img_size, 3))
    seg_map_louti = SegmentationMapsOnImage(np.expand_dims(seg_arr_louti, axis=-1), shape=(img_size, img_size, 3))
     
    for i in range(10):
        # 原始图片 image 与 标注 segmentation_maps 一起变化
        # s_augs 是一个 list，包含两个 labels
 
        scale_x = get_random(0.8, 1.2)
        scale_y = get_random(0.8, 1.2)
 
        translate_percent_x = get_random(-0.2, 0.2)
        translate_percent_y = get_random(-0.2, 0.2)
 
        rotate_ = get_random(-45, 45)
 
        shear_ = get_random(-16, 16)
 
        order_ = round(random.random())
 
        sometimes_ = round(random.random())
 
        if sometimes_:
            seq = iaa.Sequential(
                [iaa.Affine(
                        scale={"x": (scale_x), "y": (scale_y)},          # 图像缩放因子
                        translate_percent={"x": (translate_percent_x), 
                                           "y": (translate_percent_y)},  # 平移比例
                        rotate=(rotate_),                                # 平移角度
                        shear=(shear_),                                  # 错切的程度
                        order=[order_],                                  # 插值顺序
                        cval=(0),                                        # 没有值的地方填充的像素点值
                        mode="constant")                                 # 填充模式，常量填充
                ])
        else:
            seq = iaa.Sequential([])
 
        i_aug, s_aug1 = seq(image = img_arr, 
                                   segmentation_maps = seg_map_louding)
        i_aug, s_aug2 = seq(image = img_arr, 
                                   segmentation_maps = seg_map_louti)        
 
        dst_img_path = os.path.join(dst_img_dir, "{}-{}.png".format(file.split(".")[0], i))
        dst_seg_path_louding = os.path.join(dst_seg_dir_louding, "{}-{}.png".format(file.split(".")[0], i))
        dst_seg_path_louti = os.path.join(dst_seg_dir_louti, "{}-{}.png".format(file.split(".")[0], i))
 
        Image.fromarray(i_aug).save(dst_img_path) 
        # 需要将得到的 s_aug.get_arr() 是 (512, 512, 1)
        # 需要转换为 (512, 512) 来显示
        Image.fromarray(s_aug1.get_arr().reshape((img_size, img_size))).save(dst_seg_path_louding) 
        Image.fromarray(s_aug2.get_arr().reshape((img_size, img_size))).save(dst_seg_path_louti) 