import os
import shutil
from PIL import Image
import numpy as np
def strong_dataset(im_path, im_name, save_path=''):
im = Image.open(im_path)
# 旋转
part1_im = im.transpose(Image.ROTATE_90)
part1_im.save(os.path.join(save_path, '90_'+im_name))
part2_im = im.transpose(Image.ROTATE_180)
part2_im.save(os.path.join(save_path, '180_'+im_name))
part3_im = im.transpose(Image.ROTATE_270)
part3_im.save(os.path.join(save_path, '270_'+im_name))
# 上下左右平移
width, height = im.size
shift_left_im = im.crop((0, 0, width-10, height))
shift_left_im.save(os.path.join(save_path, 'shift_left_'+im_name))
shift_right_im = im.crop((10, 0, width, height))
shift_right_im.save(os.path.join(save_path, 'shift_right_'+im_name))
shift_up_im = im.crop((0, 0, width, height-10))
shift_up_im.save(os.path.join(save_path, 'shift_up_'+im_name))
shift_down_im = im.crop((0, 10, width, height))
shift_down_im.save(os.path.join(save_path, 'shift_down_'+im_name))
# 缩小
small_im = im.resize((int(width*0.8), int(height*0.8)))
small_im.save(os.path.join(save_path, 'small_'+im_name))
# 水平翻转
flip_im = im.transpose(Image.FLIP_LEFT_RIGHT)
flip_im.save(os.path.join(save_path, 'flip_'+im_name))
# 增加图像噪声
noisy_im = np.array(im)
noise = np.random.normal(0, 25, noisy_im.shape)
noisy_im = noisy_im + noise
noisy_im = np.clip(noisy_im, 0, 255)
noisy_im = Image.fromarray(noisy_im.astype('uint8'))
noisy_im.save(os.path.join(save_path, 'noisy_'+im_name))
# 增亮
bright_im = ImageEnhance.Brightness(im).enhance(1.5)
bright_im.save(os.path.join(save_path, 'bright_'+im_name))
if __name__ == '__main__':
# 遍历一个文件夹所有的图片进行处理,并保存
ori_path = r'VOCdevkit\VOC2007\JPEGImages'
for ori_name in os.listdir(ori_path):
temp = os.path.join(ori_path, ori_name)
strong_dataset(temp, ori_name, save_path=ori_path)
前言
参加了“智蔗见智·向新而生”广西第二届人工智能大赛,但是他给的数据集只有不到100张,对于语义分割模型的训练来说实在是杯水车薪,我们队的处理方案是:先按照十字架方式裁剪图片,之后再进行3个角度的翻转,一系列数据增强的操作了数据集的数量是原来的4*4=16倍,已经可以满足模型的训练要求!而本文正是记录旋转这个数据增强的方式。
记录按十字架裁剪图片方式如下链接:
【数据增强】90°、180°和270°翻转图片(*4)
效果展示
旋转之前的图片:
旋转之后的图片:
90°
180°
270°
成品库
'''
@param:im_path,图片路径,例如:img/test.png
@param:im_name,图片名称,例如:test.png
@param:save_path,保存路径,默认为工作路径
'''
import os
import shutil
from PIL import Image
'''
@param:im_path,图片路径,例如:img/test.png
@param:im_name,图片名称,例如:test.png
@param:save_path,保存路径,默认为工作路径
'''
def strong_dataset(im_path, im_name, save_path=''):
im = Image.open(im_path)
part1_im = im.transpose(Image.ROTATE_90)
part1_im.save(os.path.join(save_path, '90_'+im_name))
part2_im = im.transpose(Image.ROTATE_180)
part2_im.save(os.path.join(save_path, '180_'+im_name))
part3_im = im.transpose(Image.ROTATE_270)
part3_im.save(os.path.join(save_path, '270_'+im_name))
1.单张图片处理代码:
import os
import shutil
from PIL import Image
def strong_dataset(im_path, im_name, save_path=''):
im = Image.open(im_path)
part1_im = im.transpose(Image.ROTATE_90)
part1_im.save(os.path.join(save_path, '90_'+im_name))
part2_im = im.transpose(Image.ROTATE_180)
part2_im.save(os.path.join(save_path, '180_'+im_name))
part3_im = im.transpose(Image.ROTATE_270)
part3_im.save(os.path.join(save_path, '270_'+im_name))
if __name__ == "__main__":
strong_dataset('img/test.png', 'test.png', save_path='img')
2.遍历目录图片代码:
import os
import shutil
from PIL import Image
def strong_dataset(im_path, im_name, save_path=''):
im = Image.open(im_path)
part1_im = im.transpose(Image.ROTATE_90)
part1_im.save(os.path.join(save_path, '90_'+im_name))
part2_im = im.transpose(Image.ROTATE_180)
part2_im.save(os.path.join(save_path, '180_'+im_name))
part3_im = im.transpose(Image.ROTATE_270)
part3_im.save(os.path.join(save_path, '270_'+im_name))
if __name__ == '__main__':
# 遍历一个文件夹所有的图片进行旋转,并且保存
ori_path = r'VOCdevkit\VOC2007\JPEGImages'
for ori_name in os.listdir(ori_path):
temp = os.path.join(ori_path, ori_name)
strong_dataset(temp, ori_name, save_path=ori_path)
3.输出说明
'''
假设:你有这样一张图片 test.png
那么裁剪后将保存为90_test.png
180_test.png
270_test.png
原图片不会删除!
'''
完毕!
如果大家觉得有用,欢迎三连~