其实，本篇应是深度学习常用图像数据增强库albumentations系列教程（三）的，但是鉴于不如现在的题目直观，还是修改了，原来两篇见如下：
深度学习常用图像数据增强库albumentations系列教程（一）
深度学习常用图像数据增强库albumentations系列教程（二）
本篇是在前面两篇基础上，对目标检测任务中常用的包围框标注数据进行增强。
1. 目标检测任务包围框
目标检测任务中在训练之前要对图像中的目标物体进行标注，比如使用labelimg对目标物体的位置和类别进行标注，生成xml文件(数据是pascal_voc格式)。
albumentations支持四种数据格式: pascal_voc，albumentations, coco和yolo，这四种数据格式使用不同的方法表示包围框的位置。
- • pascal_voc: 使用[x_min, y_min, x_max, y_max]描述包围框。x_min和y_min是包围框左上角的坐标，y_min和y_max是右下角的坐标,如[138, 103, 161, 471]
- • albumentations: 使用[x_min, y_min, x_max, y_max]表示，和pascal_voc不同的是albumentations用的是归一化的值去描述，即将横纵坐标除以相应的长宽，如[138/640, 103/480, 161/640, 471/480]
- • coco: 使用[x_min, y_min, width, height]表示包围框，如[138, 103, 23, 368]
- • yolo: 使用[x_center, y_center, width, height]，前面两个参数是规范化后的包围框的中心位置，如[((138 161)/2)/640, ((103 471)/2)/480, 23/640, 368/480]
2. 目标检测任务图像-标注对数据增强功能实现
针对训练样本量少的情况，我们常常会使用数据增强的方法增加样本量，如图像的旋转、平移、缩放、改变亮度等，针对增强后的图像常常还需要标注，标注工作量较大。尽管有些方法在训练的时候会帮你实现这些功能，我个人还是习惯将标注增强直观展示，确定标注增强的合理性。
图像-标注对增强包括如下流程：
- 1. 利用单张或者多张图像进行标注，生成xml文件
- 2. 定义增强pipeline
- 3. 从文件夹中遍历原始的图像文件和xml文件
- 4. 通过增强pipeline得到图像标注增强对用于训练
代码语言：javascript复制注意：不是所有的变换都支持包围框标注数据增强的，目前（20220921）支持包围框增强的变换
import random
import cv2
from matplotlib import pyplot as plt
import xml.etree.ElementTree as ET
import albumentations as A
import os
import time
import glob
from tqdm import trange
BOX_COLOR = (255, 0, 0) # Red
TEXT_COLOR = (255, 255, 255) # White
# original pictures size：62, then total size is 62*GENERATED_PICS_SIZE
GENERATED_PICS_SIZE = 600
def visualize_bbox(img, bbox, class_name, color=BOX_COLOR, thickness=2):
"""Visualizes a single bounding box on the image"""
# x_min, y_min, w, h = bbox
# x_min, x_max, y_min, y_max = int(x_min), int(x_min w), int(y_min), int(
# y_min h)
x_min, y_min, x_max, y_max = bbox
print(x_min, y_min, x_max, y_max)
cv2.rectangle(img, (int(x_min), int(y_min)), (int(x_max), int(y_max)),
color=color, thickness=thickness)
((text_width, text_height), _) = cv2.getTextSize(class_name,
cv2.FONT_HERSHEY_SIMPLEX,
0.35, 1)
cv2.rectangle(img, (int(x_min), int(y_min) - int(1.3 * text_height)),
(int(x_min) text_width, int(y_min)), BOX_COLOR, -1)
cv2.putText(
img,
text=class_name,
org=(int(x_min), int(y_min) - int(0.3 * text_height)),
fontFace=cv2.FONT_HERSHEY_SIMPLEX,
fontScale=0.35,
color=TEXT_COLOR,
lineType=cv2.LINE_AA,
)
return img
def visualize(image, bboxes, category_ids, category_id_to_name):
img = image.copy()
for bbox, category_id in zip(bboxes, category_ids):
class_name = category_id_to_name[category_id]
img = visualize_bbox(img, bbox, class_name)
plt.axis('off')
plt.imshow(img)
plt.show()
def saveNewAnnotation(new_xml_path, new_jpg_path, xml_path, bboxes, cur_dir):
in_file = open(os.path.join(xml_path), encoding='utf-8')
new_file = in_file
tree = ET.parse(new_file)
root = tree.getroot()
root[0].text = "annotation_out"
root[1].text = new_jpg_path
root[2].text = cur_dir '\annotation_out\' new_jpg_path
idx = 0
for obj in root.iter('object'):
obj[4][0].text = str(round(bboxes[idx][0]))
obj[4][1].text = str(round(bboxes[idx][1]))
obj[4][2].text = str(round(bboxes[idx][2]))
obj[4][3].text = str(round(bboxes[idx][3]))
idx = 1
tree.write(new_xml_path, 'UTF-8')
def getAnnotation(xml_path):
'''
:param xml_path:
:return: bboxes, category_ids
'''
in_file = open(os.path.join(xml_path), encoding='utf-8')
try:
tree = ET.parse(in_file)
except:
return [], []
root = tree.getroot()
bboxes = []
category_ids = []
for obj in root.iter('object'):
cls = obj.find('name').text
xmlbox = obj.find('bndbox')
bbox = [int(float(xmlbox.find('xmin').text)),
int(float(xmlbox.find('ymin').text)),
int(float(xmlbox.find('xmax').text)),
int(float(xmlbox.find('ymax').text))]
bboxes.append(bbox)
category_ids.append(cls)
return bboxes, category_ids
def main(cur_dir):
PICS_PATH = 'annotation_ori' # 存放图片的文件夹路径
paths = glob.glob(os.path.join(PICS_PATH, '*.jpg'))
for i in trange(len(paths)):
jpg_path = paths[i]
xml_path = jpg_path.split('.')[0] ".xml"
# print(jpg_path.split('.'))
new_jpg_path_prefix = 'annotation_out\'
for i in range(GENERATED_PICS_SIZE):
image = cv2.imread(jpg_path)
# print(width, ", ", height)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
new_jpg_path = jpg_path.split('.')[0].split('\')[-1] str(
i 1).zfill(4) ".jpg"
new_xml_path = new_jpg_path_prefix
jpg_path.split('.')[0].split('\')[-1] str(
i 1).zfill(4) ".xml"
bboxes, category_ids = getAnnotation(xml_path=xml_path)
if len(bboxes) == 0 & len(category_ids) == 0:
continue
category_id_to_name = {}
for i in range(len(category_ids)):
category_id_to_name[category_ids[i]] = category_ids[i]
# 变换操作
# 水平反转，高斯模糊，gamma变换，亮度变化，
transform = A.Compose(
[
A.HorizontalFlip(p=0.5),
A.Rotate(limit=2, p=0.3),
A.ShiftScaleRotate(shift_limit=0.0625,scale_limit=0, rotate_limit=0,p=0.3),
A.GaussianBlur(blur_limit=1, p=0.5),
A.ColorJitter(brightness=0.05, contrast=0.05,
saturation=0.02,
hue=0.02, always_apply=False, p=1)
],
bbox_params=A.BboxParams(format='pascal_voc',
label_fields=['category_ids']),
)
transformed = transform(image=image, bboxes=bboxes,
category_ids=category_ids)
image = transformed['image']
bboxes = transformed['bboxes']
category_ids = transformed['category_ids']
# print(bboxes)
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
cv2.imencode('.jpg', image)[1].tofile("annotation_out\" new_jpg_path)
# visualize(image, bboxes, category_ids, category_id_to_name)
saveNewAnnotation(new_xml_path, new_jpg_path, xml_path, bboxes, cur_dir)
time.sleep(1)
if __name__ == '__main__':
import os
cur_dir = os.path.dirname(os.path.abspath(__file__)) # 上级目录
print(cur_dir)
main(cur_dir)通过上述代码，我们会生成大量基于原始图像-标注对的衍生图像-标注对。
建议先生成少量图像-标注对，然后使用labelimg查看下生成的图像-标注对是否正确，确定无问题后，再生成大量图片-标注对
