深度学习语义分割图像PIL数据增强

在做训练时，遇到数据集少，检查效果不好，可以考虑数据增强。

以下为代码：图形处理库为 PIL

import cv2
import os
import random
import numpy as np
from PIL import Image
from PIL import ImageEnhance


# 随机噪声
def add_noise(img):
    img.convert("RGB")
    img_np = np.asarray(img).copy()
    for i in range(200):
        temp_x = np.random.randint(0, img_np.shape[0])
        temp_y = np.random.randint(0, img_np.shape[1])
        img_np[temp_x][temp_y] = 255
        img = Image.fromarray(np.uint8(img_np))
    return img

def add_noise_png(img):

    return img
def lum_big(image):
    enh_bri = ImageEnhance.Brightness(image)
    new_15 = enh_bri.enhance(factor=1.4)
    new_12 = enh_bri.enhance(factor=1.2)
    new_08 = enh_bri.enhance(factor=0.8)
    new_05 = enh_bri.enhance(factor=0.6)
    return new_05, new_08, new_12, new_15
def lum_big_png(image):

    return image, image, image, image
def rotate_image(image_, angle):
    rows, cols = image_.shape[:2]
    M = cv2.getRotationMatrix2D((cols / 2, rows / 2), angle, 1)
    return cv2.warpAffine(image_, M, (cols, rows))
"""
保存图像路径
"""
p_ng=r"E:\deeplabv3dp\VOCdevkit\VOC2007\SegmentationClass"
j_pg = r"E:\deeplabv3dp\VOCdevkit\VOC2007\JPEGImages"
"""
读取图像路径
"""
dir_path = r"E:\deeplabv3dp\datasets\JPEGImages"
dir_path_1 = r"E:\deeplabv3dp\datasets\SegmentationClass"
jp_list = os.listdir(dir_path)

png_list = os.listdir(dir_path_1)
for item in range(len(jp_list)):
    image_path = os.path.join(dir_path, jp_list[item])
    # jpg
    image = Image.open(image_path)
    image_path_png = os.path.join(dir_path_1, png_list[item])
    # png
    image_png = Image.open(image_path_png)
    image_png_nose = add_noise_png(image_png)

    writ_path_1_png_nosr = os.path.join(p_ng, "nose" + png_list[item])
    image_png_nose.save(writ_path_1_png_nosr)


    writ_path_1 = os.path.join(j_pg, jp_list[item])
    image.save(writ_path_1)
    image_nose = add_noise(image)
    writ_path_nose = os.path.join(j_pg, "nose" + jp_list[item])
    image_nose.save(writ_path_nose)
    writ_path_lum = os.path.join(j_pg, "lum" + jp_list[item])
    image_list = lum_big(image)
    num_jp=0
    for i in image_list:
        writ_path_lum = os.path.join(j_pg, f"lum{num_jp}" + jp_list[item])
        i.save(writ_path_lum)
        num_jp=num_jp+1

    writ_path_1_png = os.path.join(p_ng, png_list[item])
    image_png.save(writ_path_1_png)

    image_list_1 = lum_big_png(image_png)
    num_id = 0
    for q in image_list_1:
        writ_path_1_png_lum = os.path.join(p_ng,
                                           f"lum{num_id}" + png_list[item])
        q.save(writ_path_1_png_lum)
        num_id =num_id+ 1





    flip_image_2 = image.transpose(Image.FLIP_TOP_BOTTOM)
    writ_path_2 = os.path.join(j_pg, "f2" + jp_list[item])
    flip_image_2.save(writ_path_2)
    flip_image_2_png = image_png.transpose(Image.FLIP_TOP_BOTTOM)
    writ_path_2_png = os.path.join(p_ng, "f2" + png_list[item])
    flip_image_2_png.save(writ_path_2_png)

    flip_image_1 = image.transpose(Image.FLIP_LEFT_RIGHT)
    writ_path_3 = os.path.join(j_pg, "f1" + jp_list[item])
    flip_image_1.save(writ_path_3)

    flip_image_1_png = image_png.transpose(Image.FLIP_LEFT_RIGHT)
    writ_path_3_png = os.path.join(p_ng, "f1" + png_list[item])
    flip_image_1_png.save(writ_path_3_png)

    angle = random.randint(-45, 45)
    scaled_image = image.rotate(angle)
    writ_path_4 = os.path.join(j_pg, "s" + jp_list[item])
    scaled_image.save(writ_path_4)

    scaled_image_png = image_png.rotate(angle)
    writ_path_4_png = os.path.join(p_ng, "s" + png_list[item])
    scaled_image_png.save(writ_path_4_png)