Using PIL.Image to generate image label datasets

Method 3: Use PIL.Image to generate an image label dataset

提示：提示：此处独立使用图像库pillow进行读图和显示

The basic situation of the data set describes the three ways to generate image label data sets in pytorch – the preface .
PIL installation: pip install pillow
Import application: from PIL import Image
Pillow official manual : https://pillow.readthedocs.io/en/latest/index.html
Need to 服装-关键点download the data set and leave a message.

---

Image common functions

Image package commonly used function modules :

im = Image.open(file#ReadPicture
Image.open(file).convert()#ColorTransform
Image.open(file).show() #Display
Image.open(file).save() #Save
Image.open(file).thumbnail() #CreateThumbnail
im.crop(box) #Copy Region
im.paste(region, box)#Paste Region
im.resize()and im.rotate()#Resize and Rotate
Image Filter: ImageFilter from PIL import ImageFilter
im.filter(ImageFilter.BLUR)# Image Blur
Image Blur BLUR, Edge CONTOUR, Detail DETAIL, Edge Enhancement EDGE_ENHANCE, EDGE_ENHANCE_MORE, EMBOSS, FIND_EDGES, Image Sharpening SHARPEN, image smoothing SMOOTH,SMOOTH_MORE

---

Clothing Type and Keypoint Image-Label Dataset

In this example, clothing types and keypoint image-label datasets, PIL.Imagemodules are introduced.
Dataset result display: (image, coordinates, type)

提示：image读图，制作生成器，并展示图像数组数据

code :dataset_by_PIL.py

# -*- coding: utf-8 -*-
# @Time    : 2022/5/19 15:48
# @Author  : Hyan Tan 
# @File    : dataset_by_PIL.py

# 1.输入图像预处理，统一尺寸。
# 2.真实值ground truth变形，img的shape = (c, h, w),label的shape=(x, y, 是否存在和显隐)
# 3.返回一个数据发生器，img用于给模型做输入，label与输出做损失计算。
import os
import numpy as np
import pandas as pd
import torch
from PIL import Image  # PIL库的Image包是基于python开发的数字图片处理包。
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as transforms  # 使用torchvision中的变换

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# 训练和验证集采用transform做变换
transform = transforms.Compose([
    transforms.Resize([256, 256]),  # 图片resize
    # transforms.RandomCrop(224),  # 随机裁剪224*224，但此处不能随机裁剪，因为坐标群不能及时变化
    # transforms.RandomHorizontalFlip(),    # 水平翻转，但此处不能随机裁剪，因为坐标群不能及时变化
    transforms.ToTensor(),  # 将图像转为Tensor,数据归一化了欸！img.float().div(255)
    # transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])   # 标准化,为了展示图像，先注释掉
])
# 测试集采用test_transform做变换
test_transform = transforms.Compose([
    transforms.Resize([288, 288]),  # 把图片resize为256*256
    transforms.RandomCrop(256),  # 随机裁剪224*224，测试时无标签
    transforms.RandomHorizontalFlip(),    # 水平翻转
    transforms.ToTensor(),  # 将图像转为Tensor
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])   # 标准化
])


class KeyPointsDataSet(Dataset):
    """服装关键点标记数据集"""

    def __init__(self, root_dir, image_set='train', transforms=None):
        """
        初始化数据集
        :param root_dir: 数据目录(.csv和images的根目录)
        :param image_set: train训练,val验证,test测试
        :param transforms（callable,optional）:图像变换-可选
        标签数据文件格式为csv_file: 标签csv文件(内容：图像相对地址-category类型-标签coordination坐标)
        """
        self._imgset = image_set
        self._image_paths = []  # 用于存储图片地址列表
        self._labels = []  # 图片标签坐标群
        self._cates = []  # 标签：服装类别
        self._csv_file = os.path.join(root_dir, image_set + '.csv')  # csv标签文件地址
        self.__getFileList()  # 获取数据(图像，坐标，类型)
        self._categories = ['blouse', 'outwear', 'dress', 'trousers', 'skirt', ]
        self._root_dir = root_dir
        self._transform = transforms

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

    def __getitem__(self, idx):
        img_id = self._image_paths[idx]
        img_id = os.path.join(self._root_dir, img_id)
        image = Image.open(img_id).convert('RGB')  # [3, 256, 256](通道数,高,宽）= (c, h, w)
        imgSize = image.size  # 原始图像宽高
        label = np.asfortranarray(self._labels[idx])  # (x, y, 显隐)=(宽，高，显隐性)
        category = self._categories.index(self._cates[idx])  # 0,1,2,3,4

        if self._transform:
            image = self._transform(image)  # 返回torch.Size([3, 256, 256])
        else:
            image.resize((256, 256))  # 使用resize
        afterSize = image.numpy().shape[1:]  # 缩放后图像的宽高
        # print(imgSize, afterSize)
        bi = np.array(afterSize) / np.array(imgSize)
        label[:, 0:2] = label[:, 0:2] * bi

        return image, label, category

    def __getFileList(self):
        file_info = pd.read_csv(self._csv_file)
        self._image_paths = file_info.iloc[:, 0]  # 第一列，相对地址列
        self._cates = file_info.iloc[:, 1]  # 第二列，服装类型：blouse,trousers,skirt,dress,outwear
        if self._imgset == 'train':
            landmarks = file_info.iloc[:, 2:26].values  # panda中DataFrame数据的读取。第3-25列为坐标群，共24组坐标，
            for i in range(len(landmarks)):  # 处理坐标数据84_497_1 to [84,497,1]
                label = []
                for j in range(24):
                    plot = landmarks[i][j].split('_')
                    coor = []
                    for per in plot:
                        coor.append(int(per))
                    label.append(coor)
                self._labels.append(np.concatenate(label))
            self._labels = np.array(self._labels).reshape((-1, 24, 3))
        else:
            self._labels = np.ones((len(self._image_paths), 24, 3)) * (-1)


def showImageAndCoor(img, coords):
    for coor in coords:
        if coor[2] == -1:
            pass
        else:
            img[:, coor[1]-1, coor[0]] = [255, 0, 0]  # (y,x)
            img[:, coor[1], coor[0]] = [255, 0, 0]  # 设置关键点位置坐标为红色，为了便于观察，将关键点四领域都设置为红色
            img[:, coor[1]+1, coor[0]] = [255, 0, 0]
            img[:, coor[1], coor[0]-1] = [255, 0, 0]
            img[:, coor[1], coor[0]+1] = [255, 0, 0]
    # 因为Image不能直接接受（3，256，256）多维数据，需要逐个击破
    # img = Image.fromarray(img * 255, mode='RGB')  # 所以这样写只能得到一条线，呵呵
    img0 = Image.fromarray(255 * img[0]).convert('L')
    img1 = Image.fromarray(255 * img[1]).convert('L')
    img2 = Image.fromarray(255 * img[2]).convert('L')
    img = Image.merge("RGB", [img0, img1, img2])

    img.show()


if __name__ == "__main__":
    fashionDataset = KeyPointsDataSet(root_dir=r"E:/Datasets/Fashion/Fashion AI-keypoints_24/train/",
                                      image_set="train",
                                      transforms=transform,
                                      )
    dataloader = DataLoader(dataset=fashionDataset, batch_size=4)
    for i_batch, data in enumerate(dataloader, 0):
        img, label, category = data
        img, label, category = img.numpy(), label.numpy(), category.numpy()  # 'torch.Tensor'不能直接显示，需要转换程io能处理的numpy数组格式。
        print(img.shape, label.shape, category)
        showImageAndCoor(img[0], label[0])
        break

---

Precautions:

1. The data structure of Image.open() reading image is also (c, h, w) = (channel, height, width), and the coordinate group is (width x, height y). Pay attention to the correspondence when scaling uniformly.
2. After making the dataset generator, when displaying image data, it is necessary to reconstruct the image mat to display the image. Right now:

 img0 = Image.fromarray(255 * img[0]).convert('L')
 img1 = Image.fromarray(255 * img[1]).convert('L')
 img2 = Image.fromarray(255 * img[2]).convert('L')
 img = Image.merge("RGB", [img0, img1, img2])

3. When PIL Image shows an image, it calls the system's default tool for displaying images.