深度学习之数据增强

这篇文章的目的是写写常见的数据增强模式，也会把对应的标注随着增强模式对应更新，增强模式会持续更新，从最简单的开始...

import random
import torchvision.transforms as transforms
import cv2
import torch
import os
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils
from skimage import io, transform
import xml.etree.ElementTree as ET
import torch.utils as utils


'''
这个脚本读取的是VOCdata，参考：https://github.com/Lxrd-AJ/YOLO_V1.git
'''

class CustomData(Dataset):
    # file_root 是图片的根路径
    # train_img 是train.txt的路径
    def __init__(self, file_root, train_img, image_size=(448, 448), transform=None, pair_transform=None):
        super(CustomData, self).__init__()
        self.file_root = file_root
        self.img_size = image_size
        self.transform = transform
        self.pair_transform = pair_transform
        with open(train_img, 'r') as f:
            self.img_and_target = [(os.path.join(file_root, 'JPEGImages', val.strip()+'.jpg'), os.path.join(file_root, 'Annotations', val.strip()+'.xml')) for val in f.readlines()]


    def __getitem__(self, item):
        image_path, target_path = self.img_and_target[item]
        image = cv2.imread(image_path)
        target = self.parseXml(target_path)

        return image, target

    def __len__(self):
        return len(self.img_and_target)

    def parseXml(self, xml_file):
        tree = ET.parse(xml_file)
        root = tree.getroot()
        size = root.find('size')
        w = int(size.find('width').text)
        h = int(size.find('height').text)

        cls_bbox = []
        for obj in root.iter('object'):
            cls = obj.find('name').text
            xmlbox = obj.find('bndbox')
            b = (float(xmlbox.find('xmin').text), float(xmlbox.find('ymin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymax').text))
            cls_bbox.append((cls, (w, h), b))
        return cls_bbox



class RandomVerticalFlip(object):
    def __init__(self, probability=1):
        self.p = probability

    def __call__(self, items):
        if random.random() < self.p:
            img, det = items
            img = cv2.flip(img, flipCode=0)
            update_det = []
            for cls, wh, bbox in det:
                xmin, ymin, xmax, ymax = bbox
                update_ymin = wh[1] - ymax   # 一种转换
                update_ymax = wh[1] - ymin
                update_det.append((cls, wh, (xmin, update_ymin, xmax, update_ymax)))
            return (img, update_det)
        else:
            return items


class RandomHorizontalFlip(object):
    def __init__(self, probability=1):
        self.p = probability

    def __call__(self, items):
        if random.random() < self.p:
            img, det = items
            img = cv2.flip(img, flipCode=1)
            update_det = []
            for cls, wh, bbox in det:
                xmin, ymin, xmax, ymax = bbox
                update_xmin = wh[0] - xmax
                update_xmax = wh[0] - xmin
                update_det.append((cls, wh, (update_xmin, ymin, update_xmax, ymax)))
            return (img, update_det)
        else:
            return items


def plot_bbox(img, bndbox):
    for cls, wh, bbox in bndbox:
        xmin, ymin, xmax, ymax = [int(val) for val in bbox]
        point_color = (random.randint(0,255), random.randint(0,255), random.randint(0,255))
        thickness = 2
        line_type = 4
        pt1, pt2 = (xmin, ymin), (xmax, ymax)
        cv2.putText(img, cls, (xmin, ymin), cv2.FONT_HERSHEY_COMPLEX, fontScale=0.5, color=point_color, thickness=thickness)
        cv2.rectangle(img, pt1, pt2, point_color, thickness, line_type)
    return img


if __name__ == '__main__':
    file_root = r'D:\data\voc\VOCdevkit\VOC2007'
    train_txt = r'D:\data\voc\VOCdevkit\VOC2007\ImageSets\Main\train.txt'
    custom_data = CustomData(file_root, train_txt)
    # data_loader = utils.data.DataLoader(customData, batch_size=1, shuffle=True, num_workers=4)
    random_vertical = RandomVerticalFlip()
    random_horizontal = RandomHorizontalFlip()
    for img, bbox in iter(custom_data):
        vertical_img = img.copy()
        vertical_bbox = bbox.copy()

        horizontal_img = img.copy()
        horizontal_bbox = bbox.copy()

        update_img = plot_bbox(img, bbox)
        # 展示原图
        cv2.imshow('update_img', update_img)
        cv2.waitKey(0)

        # 展示水平翻转
        vertical_img, vertical_bbox = random_vertical((vertical_img, vertical_bbox))
        vertical_update_img = plot_bbox(vertical_img, vertical_bbox)
        cv2.imshow('vertical_update_img', vertical_update_img)
        cv2.waitKey(0)

        # 展示垂直翻转
        horizontal_img, horizontal_bbox = random_horizontal((horizontal_img, horizontal_bbox))
        horizontal_update_img = plot_bbox(horizontal_img, horizontal_bbox)
        cv2.imshow('horizontal_update_img', horizontal_update_img)
        cv2.waitKey(0)