目录
前言:
示例项目数据结构:
源代码:
运行代码后生成的项目结构:
效果:
前言:
使用yolo训练模型时,遇到数据集很小的情况(一两百张),训练出来的模型效果不好,可以选择扩充数据集重新训练,这篇文章提供将照片左右翻转,上下翻转,以及将标注信息进行对应的翻转,这里的标注信息是yolo五列格式的:
分别代表:类别、归一化的中心点的x坐标、归一化的中心点的y坐标、归一化的目标狂的宽、归一化的目标框的高
示例项目数据结构:
源代码:
from PIL import Image
import os
root_dir = 'D:\python\PycharmProjects\photo-deal\\transpose-image\data'
images_dir = "images"
labels_dir = "labels"
for file in os.listdir(os.path.join(root_dir, images_dir)):
name, ext = os.path.splitext(file)
image = Image.open(os.path.join(root_dir, images_dir, file))
# 获取图像的长和宽
width, height = image.size
image.transpose(Image.FLIP_TOP_BOTTOM).save(os.path.join(root_dir, images_dir, name + '_tb' + ext)) # 上下翻转
# 读取对应的txt
tbtext_list = []
with open(os.path.join(root_dir, labels_dir, name + '.txt'), "r") as annofile:
for line in annofile:
line = line.strip()
rects = line.split(" ")
tbtext_list.append(
rects[0] + " " + rects[1] + " " + "{:.6f}".format((1.0 - float(rects[2]))) + " " + rects[3] + " " +
rects[4])
with open(os.path.join(root_dir, labels_dir, name + '_tb.txt'), 'a') as classify_annofile:
for item in tbtext_list:
classify_annofile.write(item + "\n") # 写入内容并换行
# 翻转图像并保存
image.transpose(Image.FLIP_LEFT_RIGHT).save(os.path.join(root_dir, images_dir, name + '_lr' + ext)) # 左右翻转
# 翻转txt文件
# 读取对应的txt
text_list = []
with open(os.path.join(root_dir, labels_dir, name + '.txt'), "r") as annofile:
for line in annofile:
line = line.strip()
rects = line.split(" ")
text_list.append(
rects[0] + " " + "{:.6f}".format((1.0 - float(rects[1]))) + " " + rects[2] + " " + rects[3] + " " +
rects[4])
with open(os.path.join(root_dir, labels_dir, name + '_lr.txt'), 'a') as classify_annofile:
for item in text_list:
classify_annofile.write(item + "\n") # 写入内容并换行
print(file + "图像已处理")
运行代码后生成的项目结构:
效果:
原图:
翻转后: