阿里巴巴网站建设的不足之处,注册域名邮箱,微网站设计与制作,做网站不备案因为有一批数据有点儿小#xff0c;数据质量不佳#xff0c;为了标注方便使用数据增强将数据固定在1080P#xff0c;方便标注#xff0c;
# -*- coding: UTF-8 -*-Project #xff1a;yolov5_relu_fire_smoke_v1.4
IDE #xff1a;PyCharm
Au…因为有一批数据有点儿小数据质量不佳为了标注方便使用数据增强将数据固定在1080P方便标注
# -*- coding: UTF-8 -*-Project yolov5_relu_fire_smoke_v1.4
IDE PyCharm
Author 沐枫
Date 2024/4/2 20:28添加白条做数据增强最后所有的图片尺寸固定在1080Pimport os
import multiprocessing
from concurrent import futures
from copy import deepcopyimport cv2
import numpy as np
import xml.etree.ElementTree as ET
import xml.dom.minidom as minidomdef xyxy2xywh(x):Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1top-left, xy2bottom-right:param x::return:y np.copy(x)y[..., 0] (x[..., 0] x[..., 2]) / 2 # x centery[..., 1] (x[..., 1] x[..., 3]) / 2 # y centery[..., 2] x[..., 2] - x[..., 0] # widthy[..., 3] x[..., 3] - x[..., 1] # heightreturn ydef xywh2xyxy(x: np.ndarray):Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1top-left, xy2bottom-right:param x::return:y np.copy(x)y[..., 0] x[..., 0] - x[..., 2] / 2 # top left xy[..., 1] x[..., 1] - x[..., 3] / 2 # top left yy[..., 2] x[..., 0] x[..., 2] / 2 # bottom right xy[..., 3] x[..., 1] x[..., 3] / 2 # bottom right yreturn ydef decodeVocAnnotation(voc_xml_path, class_index_dict):voc数据集格式的文件解析将一个文件解析成一个list使用空格间隔不同对象注意:返回的类别不是整型而是字符串的类别名称注意判断返回值是否为 空如果是空说明没有目标是一张背景图:param voc_xml_path: xml路径:param class_index_dict: 类别字典:return: [(cls_index, x1, y1, x2, y2), ...]assert voc_xml_path.endswith(.xml), voc_xml_path must endswith .xmlwith open(voc_xml_path, r, encodingutf-8) as xml_file:# 打开xml文件并返回根节点root ET.ElementTree().parse(xml_file)# 定义一个列表专门保存目标information []# 查找root节点下所有目标信息for obj in root.iter(object):# 目标的名称name obj.find(name).text# 目标的bbox坐标一般voc是保存的corner格式的bboxbox obj.find(bndbox)xmin box.find(xmin).textymin box.find(ymin).textxmax box.find(xmax).textymax box.find(ymax).text# 添加一个目标的信息# NOTE:返回值的listinformation.append((class_index_dict[name], int(xmin), int(ymin), int(xmax), int(ymax)))return informationdef create_voc_xml(image_folder, image_filename, width: int, height: int, labels,save_root, class_name_dict, conf_thresh_dictNone)::param image_folder: 图片的相对路径:param image_filename: 000001.jpg:param width: 图片宽:param height: 图片高:param labels: 目标框:[[class_index, xmin, ymin, xmax, ymax], ...]:param save_root: 保存xml的根目录:param class_name_dict: cls_index:cls_name,根据index获取正确的类别name:param conf_thresh_dict: cls_index:conf_thresh,根据不同类别设置的阈值获取对应的目标如果设置为None则表示保存的txt没有置信度:return:# 创建 XML 文件的根元素root ET.Element(annotation)# 添加图片信息folder ET.SubElement(root, folder)folder.text str(image_folder)# 图片名字filename ET.SubElement(root, filename)filename.text os.path.join(image_filename)# 图片大小size ET.SubElement(root, size)width_element ET.SubElement(size, width)width_element.text str(width)height_element ET.SubElement(size, height)height_element.text str(height)depth ET.SubElement(size, depth) # 通道数depth.text 3# 添加目标框信息for label in labels:# 如果该参数设置为None表示保存的txt没有Noneif conf_thresh_dict is None:# 保证这几项是整数class_index, x1, y1, x2, y2 label.astype(dtypenp.int32)else:class_index, x1, y1, x2, y2, conf label# 保证这几项是整数class_index, x1, y1, x2, y2 np.array([class_index, x1, y1, x2, y2], dtypenp.int32)# 根据置信度过滤是否保存项if conf conf_thresh_dict[class_index]:continueobj ET.SubElement(root, object)name ET.SubElement(obj, name)name.text class_name_dict[int(class_index)]pose ET.SubElement(obj, pose)pose.text Unspecifiedtruncated ET.SubElement(obj, truncated)truncated.text 0difficult ET.SubElement(obj, difficult)difficult.text 0bndbox ET.SubElement(obj, bndbox)xmin ET.SubElement(bndbox, xmin)xmin.text str(x1)ymin ET.SubElement(bndbox, ymin)ymin.text str(y1)xmax ET.SubElement(bndbox, xmax)xmax.text str(x2)ymax ET.SubElement(bndbox, ymax)ymax.text str(y2)# 创建 XML 文件并保存xml_str ET.tostring(root, encodingutf-8)xml_str minidom.parseString(xml_str)# 设置缩进为4个空格xml可读性提高pretty_xml xml_str.toprettyxml(indent * 4)save_path os.path.join(save_root, f{os.path.splitext(image_filename)[0]}.xml)os.makedirs((os.path.dirname(save_path)), exist_okTrue)with open(save_path, w) as xmlFile:xmlFile.write(pretty_xml)def resize_and_pad(image: np.ndarray, labels: np.ndarray, width1920, height1080)::param image::param labels: (cls_id, x, y, w, h):param width::param height::return:def _resize(image: np.ndarray, labels: np.ndarray, width, height)::param image::param labels: (cls_id, x, y, w, h):param width::param height::return:image: 最后的图片labels: (cls_id, x, y, w, h)# 判断图片的尺寸如果尺寸比目标尺寸大就等比例缩放斌使用纯白色填充如果尺寸比目标尺寸小就直接填充到目标尺寸img_h, img_w image.shape[:2]if img_w width and img_h height:# 直接填充# 填充的宽度和高度dw (width - img_w) // 2dh (height - img_h) // 2# 创建一个新的蒙版new_image np.ones(shape(height, width, 3), dtypenp.uint8) * 255# 将图片填充到里面new_image[dh:dh img_h, dw:dw img_w, :] image[:, :, :]# 标签平移(cls_id, x, y, w, h)labels[..., 1] dwlabels[..., 2] dhelse:# 等比例缩放后再填充# 计算宽度和高度的缩放比例ratio min((width / img_w), (height / img_h))# 计算缩放后的宽度和高度new_width int(img_w * ratio)new_height int(img_h * ratio)# 等比例缩放图像resized_img cv2.resize(image, (new_width, new_height))# 计算需要填充的宽度和高度dw (width - new_width) // 2dh (height - new_height) // 2# 创建一个新的蒙版new_image np.ones(shape(height, width, 3), dtypenp.uint8) * 255# 将图片填充到里面new_image[dh:dh new_height, dw:dw new_width, :] resized_img[:, :, :]# 标签缩放平移(cls_id, x, y, w, h)labels[..., 1:] * ratio # 坐标和宽高都需要缩放# 只有中心点需要平移不影响宽高labels[..., 1] dwlabels[..., 2] dhreturn new_image, labelsSCALE 2# 原图的宽高img_h, img_w image.shape[:2]# NOTE:先在外面扩大一次写和内部函数在判断如果图片比目标尺寸大就等比例缩放如果图片比目标尺寸小就直接填充# 比较小先扩大再等比例缩放比较大直接等比例缩放if img_w width and img_h height:new_w img_w * SCALEnew_h img_h * SCALE# 图片扩大为原来的2倍image cv2.resize(image, (new_w, new_h))# labels也扩大,因为图片扩大2倍所以目标的中心点和宽高都会扩大同样的倍数labels[..., 1:] * SCALE# 缩放和填充new_image, new_labels _resize(image, labels, widthwidth, heightheight)return new_image, new_labelsdef run(image_path, xml_root,image_root, save_image_root, save_xml_root,class_index_dict, class_name_dict):image_file os.path.basename(image_path)image_name, suffix os.path.splitext(image_file)xml_path image_path.replace(image_root, xml_root).replace(suffix, .xml)if not os.path.exists(xml_path):print(f\n{image_path} no xml\n)returntry:# 读图image cv2.imread(image_path)if image is None:return# (cls_id, x, y, w, h)labels decodeVocAnnotation(xml_path, class_index_dict)if len(labels) 0:print(f\n{image_path} no label\n)returnlabels np.array(labels, dtypenp.float64)if labels.ndim 2:labels labels[None, ...]# 坐标框转成xywhlabels[..., 1:] xyxy2xywh(labels[..., 1:].copy())# resize and padnew_image, new_labels resize_and_pad(image, labels.copy(), width1920, height1080)new_img_h, new_img_w new_image.shape[:2]# 坐标框转成xyxynew_labels[..., 1:] xywh2xyxy(new_labels[..., 1:].copy())# 开始保存# save_image_path image_path.replace(image_root, save_image_root)save_image_path os.path.join(save_image_root,os.path.basename(os.path.dirname(image_path)),faug_{image_file})save_xml_path save_image_path.replace(save_image_root, save_xml_root).replace(suffix, .xml)os.makedirs(os.path.dirname(save_image_path), exist_okTrue)os.makedirs(os.path.dirname(save_xml_path), exist_okTrue)# 保存图片cv2.imwrite(save_image_path, new_image)# 创建xmlcreate_voc_xml(image_foldersave_image_path.replace(save_image_root os.sep, ),image_filenameos.path.basename(save_image_path),widthnew_img_w,heightnew_img_h,labelsnp.array(new_labels, dtypenp.int32),save_rootos.path.dirname(save_xml_path),class_name_dictclass_name_dict, )print(f\r{image_path}, end)except Exception as e:print(f{image_path} {run.__name__}:{e})def run_process(root_file_list,image_root, xml_root, save_image_root, save_xml_root,class_index_dict, class_name_dict):# 使用线程池控制程序执行with futures.ThreadPoolExecutor(max_workers5) as executor:for root, file in root_file_list:# 向线程池中提交任务向线程池中提交任务的时候是一个一个提交的image_path os.path.join(root, file)executor.submit(run,*(image_path, xml_root,image_root, save_image_root, save_xml_root,class_index_dict, class_name_dict))if __name__ __main__:class_index_dict {fire: 0,smoke: 1,}class_name_dict {0: fire,1: smoke,}data_root rZ:\Datasets\FireSmoke_v4data_root os.path.abspath(data_root)# 数据的原图根目录image_root os.path.join(data_root, images)# xml标注文件根目录xml_root os.path.join(data_root, annotations)# 保存根目录save_image_root os.path.join(image_root, aug-pad)save_xml_root os.path.join(xml_root, aug-pad)# 过滤点不想用的目录exclude_dirs [os.sep rbackground,os.sep rcandle_fire,os.sep rAUG,os.sep rsmoke,os.sep rval,os.sep raug-merge,os.sep rcandle_fire,os.sep rcut_aug,os.sep rminiFire,os.sep rnet,os.sep rnew_data,os.sep rrealScenario,os.sep rsmoke,os.sep rTSMCandle,]max_workers 10 # 线程/进程 数print(fmax_workers:{max_workers})# 一个进程处理多少图片max_file_num 3000# 保存root和file的listroot_file_list list()# 创建进程池根据自己的设备自行调整别太多否则会变慢pool multiprocessing.Pool(processesmax_workers)for root, _, files in os.walk(image_root):# 只要其中有一个就跳过if any(map(lambda x: x in root, exclude_dirs)):continuefor file in files:file_name, suffix os.path.splitext(file)if suffix.lower() not in (.jpg, .jpeg, .bmp, .png):continueroot_file_list.append((root, file))if len(root_file_list) max_file_num:pool.apply_async(run_process,(deepcopy(root_file_list),image_root, xml_root, save_image_root, save_xml_root,class_index_dict, class_name_dict))# 清除列表中的存储root_file_list.clear()else:pool.apply_async(run_process,(deepcopy(root_file_list),image_root, xml_root, save_image_root, save_xml_root,class_index_dict, class_name_dict))# 清除列表中的存储root_file_list.clear()# 关闭进程池pool.close()# 等待所有子进程执行结束pool.join()print(\nFinish ...)
