遥感影像语义分割——数据增强（图像和标签同时增强）

文章目录

遥感影像语义分割——数据增强（图像和标签同时增强）8位图像与24位图像数据增强

8位图像与24位图像

​ Labelme标注图像生成的标签图为8位彩色图，在Python中用PIL查看图片模式为**‘P’**。在深度学习做训练时， 输入的训练图像需要8位彩色图或者8位灰度图。下面演示24位与这两种模式的转换。

​ 24位彩色图转8位色彩图。在转8位彩色图之后，可能会出现像素值的变化，博主目前也没有找到原因和解决方法。如果有知道解决方法的也欢迎交流。

from PIL import Image#打开24位图像img = Image.open('000.png')#转8位彩色图new_img = img.convert('P')

​ 24位彩色图转8位灰度图。

from PIL import Image#打开24位图像img = Image.open('000.png')#转8位彩色图new_img = img.convert('L')

数据增强

​ 如果用opencv来做数据增强会导致输入的8位彩色图变成24位彩色图。这样会导致数据集出现问题无法训练。因此下面介绍使用PIL做数据增强，包括旋转角度、翻转、色度、对比度、亮度的改变。

from PIL import Image, ImageFont, ImageDraw, ImageEnhanceimport matplotlib.pyplot as pltimport numpy as npimport randomimport randomimport osdef image_rotate(image,label):"""对图像进行一定角度的旋转:param image_path: 图像路径:param save_path: 保存路径:param angle: 旋转角度:return:"""image_rotated = image.transpose(Image.ROTATE_90).convert('RGB')label_rotated = label.transpose(Image.ROTATE_90)return image_rotated,label_rotateddef image_rotate1(image,label):"""对图像进行一定角度的旋转:param image_path: 图像路径:param save_path: 保存路径:param angle: 旋转角度:return:"""image_rotated = image.transpose(Image.ROTATE_270).convert('RGB')label_rotated = label.transpose(Image.ROTATE_270)return image_rotated,label_rotateddef bright(image):enh_bri = ImageEnhance.Brightness(image)brightness = 1.2image_brightened = enh_bri.enhance(brightness)return image_brightened.convert('RGB')def ruidu(image):enh_sha = ImageEnhance.Sharpness(image)sharpness = 2.3image_sharped = enh_sha.enhance(sharpness)return image_sharped.convert('RGB')def sedu(image):enh_col = ImageEnhance.Color(image)color = 1.2image_colored = enh_col.enhance(color)return image_colored.convert('RGB')def duibidu(image):enh_con = ImageEnhance.Contrast(image)contrast = 1.3image_contrasted = enh_con.enhance(contrast)return image_contrasted.convert('RGB')def image_flip(image,label):image_transpose = image.transpose(Image.FLIP_LEFT_RIGHT).convert('RGB')label_transpose = label.transpose(Image.FLIP_LEFT_RIGHT)return image_transpose,label_transposedef image_color(image,label):image_transpose = image.transpose(Image.FLIP_TOP_BOTTOM).convert('RGB')label_transpose = label.transpose(Image.FLIP_TOP_BOTTOM)return image_transpose,label_transposepath_img = r'E:\torch-deeplabv3\pytorch-deeplab-xception-master\Waste\JPEGImages'path_label = r'E:\torch-deeplabv3\pytorch-deeplab-xception-master\Waste\SegmentationClass'path_new_img = r'E:\aa\torch-deeplabv3\pytorch-deeplab-xception-master\Waste\JPEGImages'path_new_label = r'E:\aa\torch-deeplabv3\pytorch-deeplab-xception-master\Waste\SegmentationClass'img_list = os.listdir(path_img)label_list = os.listdir(path_label)k=0for i in range(len(img_list)):img = Image.open(path_img + '/' + img_list[i])label =Image.open(path_label + '/' + img_list[i][0:-4] + '.png')#保存原图img.convert('RGB').save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')label.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k += 1#角度旋转第一次img1,mask = image_rotate(img,label)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')mask.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#角度旋转第二次img1,mask = image_rotate1(img,label)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')mask.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#调整亮度img1 = bright(img)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')label.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#调整对比度img1 = duibidu(img)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')label.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#调整锐度img1 = ruidu(img)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')label.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#调整色度img1 = sedu(img)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')label.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#左右翻转img1,mask = image_flip(img,label)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')mask.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k+=1#上下翻转img1,mask = image_color(img,label)img1.save(path_new_img + '/' + str(("%05d" % (k))) + '.jpg')mask.save(path_new_label + '/' + str(("%05d" % (k))) + '.png')k += 1print(img_list[i] + 'is finished')