大家好,欢迎来到IT知识分享网。
前言:
由于最近要写论文,但是收集到的数据集又太少无奈只能使用数据增强的方法来扩充自己的数据集,下面我将整理一下扩充该数据集的主要方法和流程。
一、什么是数据增强
数据增强是指通过使用多种方法和技术来增加训练数据的数量和多样性,以改善机器学习模型的性能和鲁棒性。在机器学习和深度学习中,数据是训练模型的关键组成部分。更多的、更多样化的数据通常有助于模型更好地泛化和适应新的数据集。
二、数据增强的类别
镜像和翻转:对图像数据进行水平或垂直翻转,创建镜像图像,增加数据多样性。
旋转和缩放:对图像进行旋转、缩放或裁剪,以增加对不同尺度和角度的变化的模型鲁棒性。
加噪声处理:向数据中添加随机噪声,如高斯噪声、随机剪裁、色彩变换等,使模型更能应对现实世界中的噪声和变化。
数据插值:通过线性或非线性插值方法来填充数据,特别适用于时间序列数据和缺失值处理。
生成对抗网络(GANs):使用生成对抗网络生成与原始数据相似但略有差异的数据,以扩展数据集。
样本合成:基于现有数据样本的属性,创造新的合成数据样本。
数据重采样:对于不平衡的数据集,增加少数类别的样本或减少多数类别的样本,以改善数据平衡。
三、如何实现数据增强
STEP1:准备工作
- img 用于存放自己手里已有的数据集图片
- img2 用于存放增强后的数据集图片
- xml 用于存放自己手里已有的数据集图片对应的标签(这里必须是VOC格式)
- xml2 用于存放增强后的数据集图片对应的标签
- txt 用于存放将xml2中的voc格式的标签转换成txt格式(yolov5识别txt格式的标签)
STEP2:使用数据增强代码
在pycharm中新建一个文件,插入以下代码
# -*- 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array bboxes: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_rate self.max_rotation_angle = max_rotation_angle self.crop_rate = crop_rate self.shift_rate = shift_rate self.change_light_rate = change_light_rate self.add_noise_rate = add_noise_rate self.flip_rate = flip_rate self.cutout_rate = cutout_rate self.cut_out_length = cut_out_length self.cut_out_holes = cut_out_holes self.cut_out_threshold = cut_out_threshold # 是否使用某种增强方式 self.is_addNoise = is_addNoise self.is_changeLight = is_changeLight self.is_cutout = is_cutout self.is_rotate_img_bbox = is_rotate_img_bbox self.is_crop_img_bboxes = is_crop_img_bboxes self.is_shift_pic_bboxes = is_shift_pic_bboxes self.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.py Randomly 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iou boxB为bouding box ''' # determine the (x, y)-coordinates of the intersection rectangle xA = 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 rectangle interArea = (xB - xA + 1) * (yB - yA + 1) # compute the area of both the prediction and ground-truth # rectangles boxAArea = (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w if img.ndim == 3: h, w, c = img.shape else: _, h, w, c = img.shape mask = 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_min y2 = np.clip(y + length // 2, 0, h) x1 = np.clip(x - length // 2, 0, w) x2 = np.clip(x + length // 2, 0, w) chongdie = False for box in bboxes: if cal_iou([x1, y1, x2, y2], box) > threshold: chongdie = True break mask[y1: y2, x1: x2, :] = 0. img = img * mask return img # ---4.旋转--- # def _rotate_img_bbox(self, img, bboxes, angle=5, scale=1.): w, h = img.shape[1], img.shape[0] rangle = np.deg2rad(angle) # angle in radians nw = (abs(np.sin(rangle) * h) + abs(np.cos(rangle) * w)) * scale nh = (abs(np.cos(rangle) * h) + abs(np.sin(rangle) * w)) * scale rot_mat = cv2.getRotationMatrix2D((nw * 0.5, nh * 0.5), angle, scale) rot_move = np.dot(rot_mat, np.array([(nw - w) * 0.5, (nh - h) * 0.5, 0])) rot_mat[0, 2] += rot_move[0] rot_mat[1, 2] += rot_move[1] rot_img = cv2.warpAffine(img, rot_mat, (int(math.ceil(nw)), int(math.ceil(nh))), flags=cv2.INTER_LANCZOS4) rot_bboxes = [] for bbox in bboxes: points = np.array([[bbox[0], bbox[1]], [bbox[2], bbox[1]], [bbox[2], bbox[3]], [bbox[0], bbox[3]]]) new_points = cv2.transform(points[None, :, :], rot_mat)[0] rx, ry, rw, rh = cv2.boundingRect(new_points) corrected_bbox = [max(0, rx), max(0, ry), min(nw, rx + rw), min(nh, ry + rh)] corrected_bbox = [int(val) for val in corrected_bbox] # Convert to int and correct order if necessary rot_bboxes.append(corrected_bbox) 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array 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 输出: crop_img:裁剪后的图像array crop_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 = 0 y_min = h y_max = 0 for 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): h, w = img.shape[:2] x = random.uniform(-w * 0.2, w * 0.2) y = random.uniform(-h * 0.2, h * 0.2) M = np.float32([[1, 0, x], [0, 1, y]]) shift_img = cv2.warpAffine(img, M, (w, h)) shift_bboxes = [] for bbox in bboxes: new_bbox = [bbox[0] + x, bbox[1] + y, bbox[2] + x, bbox[3] + y] corrected_bbox = [max(0, new_bbox[0]), max(0, new_bbox[1]), min(w, new_bbox[2]), min(h, new_bbox[3])] corrected_bbox = [int(val) for val in corrected_bbox] # Convert to int and correct order if necessary shift_bboxes.append(corrected_bbox) return shift_img, shift_bboxes # ---7.镜像--- # def _filp_pic_bboxes(self, img, bboxes): # Randomly decide the flip method flipCode = random.choice([-1, 0, 1]) # -1: both; 0: vertical; 1: horizontal flip_img = cv2.flip(img, flipCode) # Apply the flip h, w, _ = img.shape flip_bboxes = [] for bbox in bboxes: x_min, y_min, x_max, y_max = bbox if flipCode == 0: # Vertical flip new_bbox = [x_min, h - y_max, x_max, h - y_min] elif flipCode == 1: # Horizontal flip new_bbox = [w - x_max, y_min, w - x_min, y_max] else: # Both flips new_bbox = [w - x_max, h - y_max, w - x_min, h - y_min] flip_bboxes.append(new_bbox) 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array bboxes: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 += 1 angle = 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 += 1 img, bboxes = self._shift_pic_bboxes(img, bboxes) if self.is_changeLight: if random.random() > self.change_light_rate: # 改变亮度 change_num += 1 img = self._changeLight(img) if self.is_addNoise: if random.random() < self.add_noise_rate: # 加噪声 change_num += 1 img = self._addNoise(img) if self.is_cutout: if random.random() < self.cutout_rate: # cutout change_num += 1 img = 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 += 1 img, 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').text box = 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='D:/Python/1.Python/Pycharm2021/yolov5-master/data_augmention/img') parser.add_argument('--source_xml_path', type=str, default='D:/Python/1.Python/Pycharm2021/yolov5-master/data_augmention/xml') parser.add_argument('--save_img_path', type=str, default='D:/Python/1.Python/Pycharm2021/yolov5-master/data_augmention/img2') parser.add_argument('--save_xml_path', type=str, default='D:/Python/1.Python/Pycharm2021/yolov5-master/data_augmention/xml2') 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 = 0 pic_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
import os import xml.etree.ElementTree as ET # xml文件存放目录(不要以\结尾) input_dir = r'D:/xxx/Data_Count/class/6.28data/6.28data' # 输出txt文件目录(不要以\结尾) out_dir = r'D:/xxx/Data_Count/class/6.28data/labels' #类别名 class_list = ['vehicle','non_vehicle'] # 获取目录所有xml文件 def file_name(input_dir): F = [] for root, dirs, files in os.walk(input_dir): for file in files: # print file.decode('gbk') #文件名中有中文字符时转码 if os.path.splitext(file)[1] == '.xml': t = os.path.splitext(file)[0] F.append(t) # 将所有的文件名添加到L列表中 return F # 返回L列表 # 获取所有分类 def get_class(filelist): for i in filelist: f_dir = input_dir + "\\" + i + ".xml" in_file = open(f_dir, encoding='UTF-8') filetree = ET.parse(in_file) in_file.close() root = filetree.getroot() for obj in root.iter('object'): difficult = obj.find('difficult').text cls = obj.find('name').text if cls not in class_list or int(difficult) == 1: class_list.append(cls) def ConverCoordinate(imgshape, bbox): # 将xml像素坐标转换为txt归一化后的坐标 xmin, xmax, ymin, ymax = bbox width = imgshape[0] height = imgshape[1] dw = 1. / width dh = 1. / height x = (xmin + xmax) / 2.0 y = (ymin + ymax) / 2.0 w = xmax - xmin h = ymax - ymin # 归一化 x = x * dw y = y * dh w = w * dw h = h * dh return x, y, w, h def readxml(i): f_dir = input_dir + "\\" + i + ".xml" txtresult = '' outfile = open(f_dir, encoding='UTF-8') filetree = ET.parse(outfile) outfile.close() root = filetree.getroot() # 获取图片大小 size = root.find('size') width = int(size.find('width').text) height = int(size.find('height').text) imgshape = (width, height) # 转化为yolov5的格式 for obj in root.findall('object'): # 获取类别名 obj_name = obj.find('name').text obj_id = class_list.index(obj_name) # 获取每个obj的bbox框的左上和右下坐标 bbox = obj.find('bndbox') xmin = float(bbox.find('xmin').text) xmax = float(bbox.find('xmax').text) ymin = float(bbox.find('ymin').text) ymax = float(bbox.find('ymax').text) bbox_coor = (xmin, xmax, ymin, ymax) x, y, w, h = ConverCoordinate(imgshape, bbox_coor) txt = '{} {} {} {} {}\n'.format(obj_id, x, y, w, h) txtresult = txtresult + txt # print(txtresult) f = open(out_dir + "\\" + i + ".txt", 'w+') f.write(txtresult) f.close() # 获取文件夹下的所有文件 filelist = file_name(input_dir) # 获取所有分类 get_class(filelist) # 打印class print(class_list) # xml转txt for i in filelist: readxml(i) # 在out_dir下生成一个class文件 f = open(out_dir + "\\classes.txt", 'a') classresult = '' for i in class_list: classresult = classresult + i + "\n" f.write(classresult) f.close()
OK 也不知道我写清楚了没有
免责声明:本站所有文章内容,图片,视频等均是来源于用户投稿和互联网及文摘转载整编而成,不代表本站观点,不承担相关法律责任。其著作权各归其原作者或其出版社所有。如发现本站有涉嫌抄袭侵权/违法违规的内容,侵犯到您的权益,请在线联系站长,一经查实,本站将立刻删除。 本文来自网络,若有侵权,请联系删除,如若转载,请注明出处:https://haidsoft.com/133808.html