yolo系列的模型在目标检测领域里面受众非常广,也十分流行,但是在使用yolo进行目标检测训练的时候,往往要将VOC格式的数据集转化为yolo专属的数据集,而yolo的训练数据集制作方法呢,最常见的也是有两种,下面我们只讲述一种最常用的方法,也是我最常使用的。
1. voc转yolo格式
我最常使用的目标检测数据集为VOC格式,而它的格式一般如下所示:
- dataset|- annotations| |- image1.xml| |- image2.xml| |- ...||- images| |- image1.jpg| |- image2.jpg| |- ...
dataset
是数据集的根目录。annotations
目录包含每个图像对应的 XML 注释文件。images
目录包含每个图像文件。
而我们要转换的yolo格式如下所示:
- dataset|- images| |- image1.jpg| |- image2.jpg| |- ...||- labels| |- image1.txt| |- image2.txt| |- ...
dataset
是数据集的根目录。images
目录包含每个图像文件,通常是以 .jpg 或 .png 等格式保存的图像文件。labels
目录包含每个图像对应的标签文件,通常是以 .txt 格式保存的文本文件。
而 labels 里面的内容填写格式为下图所示:
通常,每行的格式为:class x_center y_center width height,其中class
代表的是图片中目标所对应的类别,x_center
, y_center
是边界框的中心点坐标相对于图像宽度和高度的归一化值,width
和 height
是边界框的宽度和高度相对于图像宽度和高度的归一化值。
举例如下:
转换代码:
import xml.etree.ElementTree as ET
import pickle
import os
from os import listdir, getcwd
from os.path import joindef convert(size, box):x_center = (box[0] + box[1]) / 2.0y_center = (box[2] + box[3]) / 2.0x = x_center / size[0]y = y_center / size[1]w = (box[1] - box[0]) / size[0]h = (box[3] - box[2]) / size[1]return (x, y, w, h)def convert_annotation(xml_files_path, save_txt_files_path, classes):xml_files = os.listdir(xml_files_path)for xml_name in xml_files:xml_file = os.path.join(xml_files_path, xml_name)out_txt_path = os.path.join(save_txt_files_path, xml_name.split('.')[0] + '.txt')out_txt_f = open(out_txt_path, 'w')tree = ET.parse(xml_file)root = tree.getroot()size = root.find('size')w = int(size.find('width').text)h = int(size.find('height').text)for obj in root.iter('object'):#difficult = obj.find('difficult').textcls = obj.find('name').text#if cls not in classes or int(difficult) == 1:#continuecls_id = classes.index(cls)xmlbox = obj.find('bndbox')b = (float(xmlbox.find('xmin').text), float(xmlbox.find('xmax').text), float(xmlbox.find('ymin').text),float(xmlbox.find('ymax').text))# b=(xmin, xmax, ymin, ymax)# print(w, h, b)bb = convert((w, h), b)out_txt_f.write(str(cls_id) + " " + " ".join([str(a) for a in bb]) + '\n')if __name__ == "__main__":# 把forklift_pallet的voc的xml标签文件转化为yolo的txt标签文件# 1、需要转化的类别,这里我直接用数字代表类别,由于我是八类,所以从0到7classes = ['0', '1', '2', '3', '4', '5', '6', '7']# 2、voc格式的xml标签文件路径xml_files1 = 'annotations'# 3、转化为yolo格式的txt标签文件存储路径save_txt_files1 = 'labels'convert_annotation(xml_files1, save_txt_files1, classes)
上面代码中注释了一部分内容,比如difficult这一项,由于我xml文件里面没有difficult,所以就注释掉了,大家按照自己的需求进行使用即可。
划分数据集
在我们进行yolo目标检测模型训练之前,需要先将数据集进行合理的划分,比如说划分为训练集:验证集=8:2,或者训练集:验证集:测试集=7:2:1。不过我一般习惯只划分训练集和验证集,也就是按8:2的比例进行划分,代码如下所示:
import os
import shutil
import random# 定义数据集文件夹路径
dataset_path = 'dataset'
images_path = os.path.join(dataset_path, 'images')
labels_path = os.path.join(dataset_path, 'labels')# 定义划分后的文件夹路径
new_path = 'mydata'
train_path = os.path.join(new_path, 'train')
val_path = os.path.join(new_path, 'val')# 创建train和val文件夹
os.makedirs(os.path.join(train_path, 'images'), exist_ok=True)
os.makedirs(os.path.join(train_path, 'labels'), exist_ok=True)
os.makedirs(os.path.join(val_path, 'images'), exist_ok=True)
os.makedirs(os.path.join(val_path, 'labels'), exist_ok=True)# 获取所有图片文件的文件名
image_files = os.listdir(images_path)
# 随机打乱文件顺序
random.shuffle(image_files)# 定义验证集所占比例
val_split = 0.1
# 计算验证集大小
num_val = int(len(image_files) * val_split)# 将数据集按照比例划分到train和val文件夹中
for i, image_file in enumerate(image_files):src_image = os.path.join(images_path, image_file)src_label = os.path.join(labels_path, image_file.replace('.jpg', '.txt'))if i < num_val:dst_image = os.path.join(val_path, 'images', image_file)dst_label = os.path.join(val_path, 'labels', image_file.replace('.jpg', '.txt'))else:dst_image = os.path.join(train_path, 'images', image_file)dst_label = os.path.join(train_path, 'labels', image_file.replace('.jpg', '.txt'))shutil.copy(src_image, dst_image)shutil.copy(src_label, dst_label)
划分完成以后的文件夹格式为:
- mydata|- train| |- images| |- labels||- val| |- images| |- labels
images和labels分别是对应的数据集图片和txt标签。
数据增强
在我们参加一些目标检测类比赛的时候,往往会遇见比赛训练集不足的情况,这将极大程度上影响我们的模型精度,这时候可能就需要用到一些数据增强方法,如翻转、随机裁剪等等。当然,yolo系列的模型一般都自带有数据增强,但是我们也可以尝试训练前进行增强看看效果。
代码如下:
# -*- coding=utf-8 -*-import time
import random
import copy
import cv2
import os
import math
import numpy as np
from skimage.util import random_noise
from lxml import etree, objectify
import xml.etree.ElementTree as ET
import argparse# 显示图片
def show_pic(img, bboxes=None):'''输入:img:图像arraybboxes:图像的所有boudning box list, 格式为[[x_min, y_min, x_max, y_max]....]names:每个box对应的名称'''for i in range(len(bboxes)):bbox = bboxes[i]x_min = bbox[0]y_min = bbox[1]x_max = bbox[2]y_max = bbox[3]cv2.rectangle(img, (int(x_min), int(y_min)), (int(x_max), int(y_max)), (0, 255, 0), 3)cv2.namedWindow('pic', 0) # 1表示原图cv2.moveWindow('pic', 0, 0)cv2.resizeWindow('pic', 1200, 800) # 可视化的图片大小cv2.imshow('pic', img)cv2.waitKey(0)cv2.destroyAllWindows()# 图像均为cv2读取
class DataAugmentForObjectDetection():def __init__(self, rotation_rate=0.5, max_rotation_angle=5,crop_rate=0.5, shift_rate=0.5, change_light_rate=0.5,add_noise_rate=0.5, flip_rate=0.5,cutout_rate=0.5, cut_out_length=50, cut_out_holes=1, cut_out_threshold=0.5,is_addNoise=True, is_changeLight=True, is_cutout=True, is_rotate_img_bbox=True,is_crop_img_bboxes=True, is_shift_pic_bboxes=True, is_filp_pic_bboxes=True):# 配置各个操作的属性self.rotation_rate = rotation_rateself.max_rotation_angle = max_rotation_angleself.crop_rate = crop_rateself.shift_rate = shift_rateself.change_light_rate = change_light_rateself.add_noise_rate = add_noise_rateself.flip_rate = flip_rateself.cutout_rate = cutout_rateself.cut_out_length = cut_out_lengthself.cut_out_holes = cut_out_holesself.cut_out_threshold = cut_out_threshold# 是否使用某种增强方式self.is_addNoise = is_addNoiseself.is_changeLight = is_changeLightself.is_cutout = is_cutoutself.is_rotate_img_bbox = is_rotate_img_bboxself.is_crop_img_bboxes = is_crop_img_bboxesself.is_shift_pic_bboxes = is_shift_pic_bboxesself.is_filp_pic_bboxes = is_filp_pic_bboxes# ----1.加噪声---- #def _addNoise(self, img):'''输入:img:图像array输出:加噪声后的图像array,由于输出的像素是在[0,1]之间,所以得乘以255'''# return cv2.GaussianBlur(img, (11, 11), 0)return random_noise(img, mode='gaussian', seed=int(time.time()), clip=True) * 255# ---2.调整亮度--- #def _changeLight(self, img):alpha = random.uniform(0.35, 1)blank = np.zeros(img.shape, img.dtype)return cv2.addWeighted(img, alpha, blank, 1 - alpha, 0)# ---3.cutout--- #def _cutout(self, img, bboxes, length=100, n_holes=1, threshold=0.5):'''原版本:https://github.com/uoguelph-mlrg/Cutout/blob/master/util/cutout.pyRandomly mask out one or more patches from an image.Args:img : a 3D numpy array,(h,w,c)bboxes : 框的坐标n_holes (int): Number of patches to cut out of each image.length (int): The length (in pixels) of each square patch.'''def cal_iou(boxA, boxB):'''boxA, boxB为两个框,返回iouboxB为bouding box'''# determine the (x, y)-coordinates of the intersection rectanglexA = max(boxA[0], boxB[0])yA = max(boxA[1], boxB[1])xB = min(boxA[2], boxB[2])yB = min(boxA[3], boxB[3])if xB <= xA or yB <= yA:return 0.0# compute the area of intersection rectangleinterArea = (xB - xA + 1) * (yB - yA + 1)# compute the area of both the prediction and ground-truth# rectanglesboxAArea = (boxA[2] - boxA[0] + 1) * (boxA[3] - boxA[1] + 1)boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3] - boxB[1] + 1)iou = interArea / float(boxBArea)return iou# 得到h和wif img.ndim == 3:h, w, c = img.shapeelse:_, h, w, c = img.shapemask = np.ones((h, w, c), np.float32)for n in range(n_holes):chongdie = True # 看切割的区域是否与box重叠太多while chongdie:y = np.random.randint(h)x = np.random.randint(w)y1 = np.clip(y - length // 2, 0,h) # numpy.clip(a, a_min, a_max, out=None), clip这个函数将将数组中的元素限制在a_min, a_max之间,大于a_max的就使得它等于 a_max,小于a_min,的就使得它等于a_miny2 = np.clip(y + length // 2, 0, h)x1 = np.clip(x - length // 2, 0, w)x2 = np.clip(x + length // 2, 0, w)chongdie = Falsefor box in bboxes:if cal_iou([x1, y1, x2, y2], box) > threshold:chongdie = Truebreakmask[y1: y2, x1: x2, :] = 0.img = img * maskreturn img# ---4.旋转--- #def _rotate_img_bbox(self, img, bboxes, angle=5, scale=1.):'''参考:https://blog.csdn.net/u014540717/article/details/53301195crop_rate输入:img:图像array,(h,w,c)bboxes:该图像包含的所有boundingboxs,一个list,每个元素为[x_min, y_min, x_max, y_max],要确保是数值angle:旋转角度scale:默认1输出:rot_img:旋转后的图像arrayrot_bboxes:旋转后的boundingbox坐标list'''# 旋转图像w = img.shape[1]h = img.shape[0]# 角度变弧度rangle = np.deg2rad(angle) # angle in radians# now calculate new image width and heightnw = (abs(np.sin(rangle) * h) + abs(np.cos(rangle) * w)) * scalenh = (abs(np.cos(rangle) * h) + abs(np.sin(rangle) * w)) * scale# ask OpenCV for the rotation matrixrot_mat = cv2.getRotationMatrix2D((nw * 0.5, nh * 0.5), angle, scale)# calculate the move from the old center to the new center combined# with the rotationrot_move = np.dot(rot_mat, np.array([(nw - w) * 0.5, (nh - h) * 0.5, 0]))# the move only affects the translation, so update the translationrot_mat[0, 2] += rot_move[0]rot_mat[1, 2] += rot_move[1]# 仿射变换rot_img = cv2.warpAffine(img, rot_mat, (int(math.ceil(nw)), int(math.ceil(nh))), flags=cv2.INTER_LANCZOS4)# 矫正bbox坐标# rot_mat是最终的旋转矩阵# 获取原始bbox的四个中点,然后将这四个点转换到旋转后的坐标系下rot_bboxes = list()for bbox in bboxes:xmin = bbox[0]ymin = bbox[1]xmax = bbox[2]ymax = bbox[3]point1 = np.dot(rot_mat, np.array([(xmin + xmax) / 2, ymin, 1]))point2 = np.dot(rot_mat, np.array([xmax, (ymin + ymax) / 2, 1]))point3 = np.dot(rot_mat, np.array([(xmin + xmax) / 2, ymax, 1]))point4 = np.dot(rot_mat, np.array([xmin, (ymin + ymax) / 2, 1]))# 合并np.arrayconcat = np.vstack((point1, point2, point3, point4))# 改变array类型concat = concat.astype(np.int32)# 得到旋转后的坐标rx, ry, rw, rh = cv2.boundingRect(concat)rx_min = rxry_min = ryrx_max = rx + rwry_max = ry + rh# 加入list中rot_bboxes.append([rx_min, ry_min, rx_max, ry_max])return rot_img, rot_bboxes# ---5.裁剪--- #def _crop_img_bboxes(self, img, bboxes):'''裁剪后的图片要包含所有的框输入:img:图像arraybboxes:该图像包含的所有boundingboxs,一个list,每个元素为[x_min, y_min, x_max, y_max],要确保是数值输出:crop_img:裁剪后的图像arraycrop_bboxes:裁剪后的bounding box的坐标list'''# 裁剪图像w = img.shape[1]h = img.shape[0]x_min = w # 裁剪后的包含所有目标框的最小的框x_max = 0y_min = hy_max = 0for bbox in bboxes:x_min = min(x_min, bbox[0])y_min = min(y_min, bbox[1])x_max = max(x_max, bbox[2])y_max = max(y_max, bbox[3])d_to_left = x_min # 包含所有目标框的最小框到左边的距离d_to_right = w - x_max # 包含所有目标框的最小框到右边的距离d_to_top = y_min # 包含所有目标框的最小框到顶端的距离d_to_bottom = h - y_max # 包含所有目标框的最小框到底部的距离# 随机扩展这个最小框crop_x_min = int(x_min - random.uniform(0, d_to_left))crop_y_min = int(y_min - random.uniform(0, d_to_top))crop_x_max = int(x_max + random.uniform(0, d_to_right))crop_y_max = int(y_max + random.uniform(0, d_to_bottom))# 随机扩展这个最小框 , 防止别裁的太小# crop_x_min = int(x_min - random.uniform(d_to_left//2, d_to_left))# crop_y_min = int(y_min - random.uniform(d_to_top//2, d_to_top))# crop_x_max = int(x_max + random.uniform(d_to_right//2, d_to_right))# crop_y_max = int(y_max + random.uniform(d_to_bottom//2, d_to_bottom))# 确保不要越界crop_x_min = max(0, crop_x_min)crop_y_min = max(0, crop_y_min)crop_x_max = min(w, crop_x_max)crop_y_max = min(h, crop_y_max)crop_img = img[crop_y_min:crop_y_max, crop_x_min:crop_x_max]# 裁剪boundingbox# 裁剪后的boundingbox坐标计算crop_bboxes = list()for bbox in bboxes:crop_bboxes.append([bbox[0] - crop_x_min, bbox[1] - crop_y_min, bbox[2] - crop_x_min, bbox[3] - crop_y_min])return crop_img, crop_bboxes# ---6.平移--- #def _shift_pic_bboxes(self, img, bboxes):'''平移后的图片要包含所有的框输入:img:图像arraybboxes:该图像包含的所有boundingboxs,一个list,每个元素为[x_min, y_min, x_max, y_max],要确保是数值输出:shift_img:平移后的图像arrayshift_bboxes:平移后的bounding box的坐标list'''# 平移图像w = img.shape[1]h = img.shape[0]x_min = w # 裁剪后的包含所有目标框的最小的框x_max = 0y_min = hy_max = 0for bbox in bboxes:x_min = min(x_min, bbox[0])y_min = min(y_min, bbox[1])x_max = max(x_max, bbox[2])y_max = max(y_max, bbox[3])d_to_left = x_min # 包含所有目标框的最大左移动距离d_to_right = w - x_max # 包含所有目标框的最大右移动距离d_to_top = y_min # 包含所有目标框的最大上移动距离d_to_bottom = h - y_max # 包含所有目标框的最大下移动距离x = random.uniform(-(d_to_left - 1) / 3, (d_to_right - 1) / 3)y = random.uniform(-(d_to_top - 1) / 3, (d_to_bottom - 1) / 3)M = np.float32([[1, 0, x], [0, 1, y]]) # x为向左或右移动的像素值,正为向右负为向左; y为向上或者向下移动的像素值,正为向下负为向上shift_img = cv2.warpAffine(img, M, (img.shape[1], img.shape[0]))# 平移boundingboxshift_bboxes = list()for bbox in bboxes:shift_bboxes.append([bbox[0] + x, bbox[1] + y, bbox[2] + x, bbox[3] + y])return shift_img, shift_bboxes# ---7.镜像--- #def _filp_pic_bboxes(self, img, bboxes):'''平移后的图片要包含所有的框输入:img:图像arraybboxes:该图像包含的所有boundingboxs,一个list,每个元素为[x_min, y_min, x_max, y_max],要确保是数值输出:flip_img:平移后的图像arrayflip_bboxes:平移后的bounding box的坐标list'''# 翻转图像flip_img = copy.deepcopy(img)h, w, _ = img.shapesed = random.random()if 0 < sed < 0.33: # 0.33的概率水平翻转,0.33的概率垂直翻转,0.33是对角反转flip_img = cv2.flip(flip_img, 0) # _flip_xinver = 0elif 0.33 < sed < 0.66:flip_img = cv2.flip(flip_img, 1) # _flip_yinver = 1else:flip_img = cv2.flip(flip_img, -1) # flip_x_yinver = -1# 调整boundingboxflip_bboxes = list()for box in bboxes:x_min = box[0]y_min = box[1]x_max = box[2]y_max = box[3]if inver == 0:# 0:垂直翻转flip_bboxes.append([x_min, h - y_max, x_max, h - y_min])elif inver == 1:# 1:水平翻转flip_bboxes.append([w - x_max, y_min, w - x_min, y_max])elif inver == -1:# -1:水平垂直翻转flip_bboxes.append([w - x_max, h - y_max, w - x_min, h - y_min])return flip_img, flip_bboxes# 图像增强方法def dataAugment(self, img, bboxes):'''图像增强输入:img:图像arraybboxes:该图像的所有框坐标输出:img:增强后的图像bboxes:增强后图片对应的box'''change_num = 0 # 改变的次数# print('------')while change_num < 1: # 默认至少有一种数据增强生效if self.is_rotate_img_bbox:if random.random() > self.rotation_rate: # 旋转change_num += 1angle = random.uniform(-self.max_rotation_angle, self.max_rotation_angle)scale = random.uniform(0.7, 0.8)img, bboxes = self._rotate_img_bbox(img, bboxes, angle, scale)if self.is_shift_pic_bboxes:if random.random() < self.shift_rate: # 平移change_num += 1img, bboxes = self._shift_pic_bboxes(img, bboxes)if self.is_changeLight:if random.random() > self.change_light_rate: # 改变亮度change_num += 1img = self._changeLight(img)if self.is_addNoise:if random.random() < self.add_noise_rate: # 加噪声change_num += 1img = self._addNoise(img)if self.is_cutout:if random.random() < self.cutout_rate: # cutoutchange_num += 1img = self._cutout(img, bboxes, length=self.cut_out_length, n_holes=self.cut_out_holes,threshold=self.cut_out_threshold)if self.is_filp_pic_bboxes:if random.random() < self.flip_rate: # 翻转change_num += 1img, bboxes = self._filp_pic_bboxes(img, bboxes)return img, bboxes# xml解析工具
class ToolHelper():# 从xml文件中提取bounding box信息, 格式为[[x_min, y_min, x_max, y_max, name]]def parse_xml(self, path):'''输入:xml_path: xml的文件路径输出:从xml文件中提取bounding box信息, 格式为[[x_min, y_min, x_max, y_max, name]]'''tree = ET.parse(path)root = tree.getroot()objs = root.findall('object')coords = list()for ix, obj in enumerate(objs):name = obj.find('name').textbox = obj.find('bndbox')x_min = int(box[0].text)y_min = int(box[1].text)x_max = int(box[2].text)y_max = int(box[3].text)coords.append([x_min, y_min, x_max, y_max, name])return coords# 保存图片结果def save_img(self, file_name, save_folder, img):cv2.imwrite(os.path.join(save_folder, file_name), img)# 保持xml结果def save_xml(self, file_name, save_folder, img_info, height, width, channel, bboxs_info):''':param file_name:文件名:param save_folder:#保存的xml文件的结果:param height:图片的信息:param width:图片的宽度:param channel:通道:return:'''folder_name, img_name = img_info # 得到图片的信息E = objectify.ElementMaker(annotate=False)anno_tree = E.annotation(E.folder(folder_name),E.filename(img_name),E.path(os.path.join(folder_name, img_name)),E.source(E.database('Unknown'),),E.size(E.width(width),E.height(height),E.depth(channel)),E.segmented(0),)labels, bboxs = bboxs_info # 得到边框和标签信息for label, box in zip(labels, bboxs):anno_tree.append(E.object(E.name(label),E.pose('Unspecified'),E.truncated('0'),E.difficult('0'),E.bndbox(E.xmin(box[0]),E.ymin(box[1]),E.xmax(box[2]),E.ymax(box[3]))))etree.ElementTree(anno_tree).write(os.path.join(save_folder, file_name), pretty_print=True)if __name__ == '__main__':need_aug_num = 5 # 每张图片需要增强的次数is_endwidth_dot = True # 文件是否以.jpg或者png结尾dataAug = DataAugmentForObjectDetection() # 数据增强工具类toolhelper = ToolHelper() # 工具# 获取相关参数parser = argparse.ArgumentParser()parser.add_argument('--source_img_path', type=str, default='images')parser.add_argument('--source_xml_path', type=str, default='Annotations')parser.add_argument('--save_img_path', type=str, default='enhance_images')parser.add_argument('--save_xml_path', type=str, default='enhance_Annotations')args = parser.parse_args()source_img_path = args.source_img_path # 图片原始位置source_xml_path = args.source_xml_path # xml的原始位置save_img_path = args.save_img_path # 图片增强结果保存文件save_xml_path = args.save_xml_path # xml增强结果保存文件# 如果保存文件夹不存在就创建if not os.path.exists(save_img_path):os.mkdir(save_img_path)if not os.path.exists(save_xml_path):os.mkdir(save_xml_path)for parent, _, files in os.walk(source_img_path):files.sort()for file in files:cnt = 0pic_path = os.path.join(parent, file)xml_path = os.path.join(source_xml_path, file[:-4] + '.xml')values = toolhelper.parse_xml(xml_path) # 解析得到box信息,格式为[[x_min,y_min,x_max,y_max,name]]coords = [v[:4] for v in values] # 得到框labels = [v[-1] for v in values] # 对象的标签# 如果图片是有后缀的if is_endwidth_dot:# 找到文件的最后名字dot_index = file.rfind('.')_file_prefix = file[:dot_index] # 文件名的前缀_file_suffix = file[dot_index:] # 文件名的后缀img = cv2.imread(pic_path)# show_pic(img, coords) # 显示原图while cnt < need_aug_num: # 继续增强auged_img, auged_bboxes = dataAug.dataAugment(img, coords)auged_bboxes_int = np.array(auged_bboxes).astype(np.int32)height, width, channel = auged_img.shape # 得到图片的属性img_name = '{}_{}{}'.format(_file_prefix, cnt + 1, _file_suffix) # 图片保存的信息toolhelper.save_img(img_name, save_img_path,auged_img) # 保存增强图片toolhelper.save_xml('{}_{}.xml'.format(_file_prefix, cnt + 1),save_xml_path, (save_img_path, img_name), height, width, channel,(labels, auged_bboxes_int)) # 保存xml文件# show_pic(auged_img, auged_bboxes) # 强化后的图print(img_name)cnt += 1 # 继续增强下一张
增强后的效果图如下所示: