Yolov5官方网络改进：增加search模块（基于迁移学习的目标检测+多模态零样本自定义标签分类网络）

1.效果展示【label：目标检测：概率+针对检测框的细分类（颜色情绪都行，此部分标签可自定义）】：

1.1目标检测+颜色识别

1.2人物检测+情绪检测

1.3针对特定类别的自动裁切效果+情绪识别结果

ps.毕业设计结束后所有代码将会开源

2.以下仅为search.py代码：

import argparse
import datetime
import os
import sys
import time
from pathlib import Path
import cv2
import torch
import torch.backends.cudnn as cudnn
from numpy import random
from Tools.AddLabels import Addlabels
from Tools.CudaIsAvailable import CudaIsAvailable
from CustomLabels import class_names_custom, class_numbers_custom
from Tools.OpenLabelWeb import OpenLabelWeb
from Tools.SaveOriImgs import save_oriImgs
from Tools.labels import class_names1, class_numbers1, class_numbers2, class_names2
from lib_forSearch.Models.experimental import attempt_load
from Tools.printTerminal import printTerminal
from lib_forSearch.Utils.datasets import LoadStreams, LoadImages
from lib_forSearch.Utils.general import check_img_size, check_imshow, non_max_suppression, scale_coords, xyxy2xywh, \strip_optimizer, set_logging, increment_path
from lib_forSearch.Utils.plots import plot_one_box
from lib_forSearch.Utils.torch_utils import select_device, time_synchronized# pip install -i https://pypi.tuna.tsinghua.edu.cn/simple some-package清华镜像# print Local Time
def LocalTime():LocalTime = time.strftime("%Y{y}%m{m}%d{d} %H{h}%M{min}%S{s}").format(y='年', m='月', d='日', h='时', min='分', s='秒')print(f'现在是北京时间:{LocalTime}')#main function
def search():# Local TimeLocalTime()# TODO printTerminalprintTerminal()# 检测目标名字与编号字典def choose():while True:classNum = input(f'请选择label编号【1:coco80、2Mechanical part defect: 、3:custom labels】：')if classNum != '':classNum = int(classNum)break# TODO CustomLabel if you want to detect custom label classes, you need 1.class_names_custom & 2.class_numbers_custom (of 3.course:custom.pt)NAMES = (class_names1, class_names2, class_names_custom)NUMBERS = (class_numbers1, class_numbers2, class_numbers_custom)classNum = classNum - 1class_names = NAMES[classNum]class_numbers = NUMBERS[classNum]print(f'System path:{sys.path}')print('\n以下是阔以被检测的类别^-^These are the number of detection objects:')for i in class_names:print('class_names[{}] is {}'.format(i, class_names[i]))return class_numbers, class_nameswhile True:class_numbers, class_names = choose()if input(f'\nAre you sure:[y/n]\n') == 'y' or '':breakprint('----------------------选一个吧Please choose one-…-搞快点come on------------------------------------')# 模型参数选择 TODO weights chooseModel_path = input('model path[such as : result/P36_v5s6_1to150/weights/best.pt ]:')if Model_path == '':Model_path = r'pretrained\yolov5s6.pt'  # 文件在项目内部，相对路径就行；如果文件在项目外，则需要绝对地址print(rf'pretrained model is :{Model_path}')# 筛选用参数Confidence = input(f'confidence:[default 0.1]:')  # 检测框的置信度（有多大把握把检测目标正好框进去），大于此的才保留if Confidence == '':Confidence = 0.1Confidence = float(Confidence)IOU = input(f'IOU:[default 0.40]:')  # 非极大抑制（IOU代表检测框的重合度），大于此的即舍弃if IOU == '':IOU = 0.4IOU = float(IOU)Img_size = 1280  # 输入图像大小，最好和网络模型对应（640或1280）print(f'Confidence:{Confidence},IOU:{IOU}')# # 是否开启二级分类器# OpenClassify = Falsewhile True:# 待检测文件地址def Input():while True:Input_path = str(input(rf'under detected file path[default :  datasets/Mac2_100/images/val]:'))if Input_path != '':return Input_pathInput_path = Input()Aclasses = ['background']  # , 'person', 'car', 'bicycle'# TODO 可以直接把想要检测的类别写入Aclasses列表，然后注释掉下面的 Addlabels() 部分。#  另外：detect（）只会检测同时包含在Aclasses和classes in model 的类别，所有放心大胆地往Aclasses中塞东西吧if input(f'all classes or specific class(es):[all/sp]') == ('sp' or ''):Addlabels(class_numbers, class_names, Aclasses)else:for i in class_names:Aclasses.append(class_names[i])print(f'classes that you choose:{Aclasses}')# 文件夹保存地址Save_path = r'output/' + Aclasses[1]# 保存文件夹名称define = '%Y_%m_%d %H_%M_%S'system_time = datetime.datetime.now().strftime(define)Save_name = system_time# 检测函数def detect(save_img=False):# TODO 阈值thresholdValue = input(f'是否交由专家标注的阈值[0.50]:')if thresholdValue == '':thresholdValue = 0.50# TODO 底线BottomLine = input(f'是否交由专家标注的底线[0.15]:')if BottomLine == '':BottomLine = 0.15print(f'thresholdValue:{thresholdValue},BottomLine:{BottomLine}')#导入opt参数source, weights, view_img, save_txt, imgsz = opt.source, opt.weights, opt.view_img, opt.save_txt, opt.img_sizesave_img = not opt.nosave and not source.endswith('.txt')  # save inference imageswebcam = source.isnumeric() or source.endswith('.txt') or source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://'))label_imgs = 0# 路径save_dir = Path(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))  # increment run(save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True)  # make direxpertPath = str(save_dir) + fr'\expertLabel{BottomLine}_{thresholdValue}'# TODO 必须放在if/else外面，不然出了if/else（即使在函数内部）也是未定义；否则用global定义全局变量# 初始化（检测设备硬件和CUDA）set_logging()device = select_device(opt.device)half = device.type != 'cpu'  # half precision only supported on CUDA# 加载模型model = attempt_load(weights, map_location=device)  # load FP32 modelstride = int(model.stride.max())  # model strideimgsz = check_img_size(imgsz, s=stride)  # check img_sizeif half:model.half()  # to FP16# # 二级分类器# classify = OpenClassify# if classify:#     modelc = load_classifier(name='resnet101', n=2)  # 初始化#     modelc.load_state_dict(#         torch.load(r'C:\Users\cleste/.cache\torch\hub\checkpoints\resnet101-5d3b4d8f.pth', map_location=device)[#             'model']).to(device).eval()# 设置Dataloadervid_path, vid_writer = None, Noneif webcam:view_img = check_imshow()cudnn.benchmark = True  # set True to speed up constant image size inferencedataset = LoadStreams(source, img_size=imgsz, stride=stride)else:dataset = LoadImages(source, img_size=imgsz, stride=stride)# 获取名称和颜色names = model.module.names if hasattr(model, 'module') else model.namescolors = [[random.randint(0, 255) for _ in range(3)] for _ in names]# 运行inferenceif device.type != 'cpu':model(torch.zeros(1, 3, imgsz, imgsz).to(device).type_as(next(model.parameters())))  # run oncet0 = time.time()for path, img, im0s, vid_cap in dataset:img = torch.from_numpy(img).to(device)img = img.half() if half else img.float()  # uint8 to fp16/32img /= 255.0  # 0 - 255 to 0.0 - 1.0if img.ndimension() == 3:img = img.unsqueeze(0)# Inferencet1 = time_synchronized()pred = model(img, augment=opt.augment)[0]# 应用NMSpred = non_max_suppression(pred, opt.conf_thres, opt.iou_thres, classes=opt.classes,agnostic=opt.agnostic_nms)t2 = time_synchronized()# # 应用分类器# if classify:#     pred = apply_classifier(pred, modelc, img, im0s)# 过程检测for i, det in enumerate(pred):# detections per imageif webcam:  # batch_size >= 1p, s, im0, frame = path[i], '%g: ' % i, im0s[i].copy(), dataset.countelse:p, s, im0, frame = path, '', im0s, getattr(dataset, 'frame', 0)p = Path(p)  # to Pathsave_path = str(save_dir / p.name)  # img.jpgtxt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}')  # img.txts += '%gx%g ' % img.shape[2:]  # print stringgn = torch.tensor(im0.shape)[[1, 0, 1, 0]]  # normalization gain whwhif len(det):# 将检测框从 img_size 缩放到 im0 sizedet[:, :4] = scale_coords(img.shape[2:], det[:, :4], im0.shape).round()# print 结果:输出信息：检测到的各个类别与个数for c in det[:, -1].unique():if names[int(c)] in Aclasses:  # 只检测Aclasses这特定的类n = (det[:, -1] == c).sum()  # 检测每个类s += f'\n-----------------------------------------------------------这是第{label_imgs + 1}个检测目标-----------------------------------------------------------'s += f"\n检查结果为   类别：{names[int(c)]}，其数量：{n} 。{'不止一个哦！' * (n > 1)} "  # 添加检测结果说明label_imgs += 1# Write 结果:# 1.保存预测信息: txt# 2.在图像上画框# 3.crop_img（裁切图片）for *xyxy, conf, cls in reversed(det):if names[int(cls)] in Aclasses:if save_img or view_img:  # Add bbox to image# TODO better to use conf(idence)label = f'{names[int(cls)]}: {conf * 100:.2f}%'plot_one_box(xyxy, im0, label=label, color=colors[int(cls)], line_thickness=3)if save_txt:  # Write to filexywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist()  # normalized xywhline = (cls, *xywh, conf) if opt.save_conf else (cls, *xywh)  # label 格式化#float格式化thresholdValue = float(thresholdValue)BottomLine = float(BottomLine)# TODO 单独保存conf从阈值到1之间的标签if (float(conf) >= thresholdValue) and (float(conf) <= 1):# passwith open(txt_path + '.txt', 'a') as f:f.write(('%g ' * len(line)).rstrip() % line + '\n')# TODO 单独保存conf从底线到阈值之间的标签if (float(conf) >= BottomLine) and (float(conf) < thresholdValue):expertLabelPath = expertPath + fr'\labels'# 不存在则创建if not os.path.exists(expertLabelPath):os.makedirs(expertLabelPath)with open(expertLabelPath + rf'\{p.stem}.txt', 'a') as f:f.write(('%g ' * len(line)).rstrip() % line + '\n')# 将检测数据统一写入包含置信度的txt为文件with open(str(save_dir) + r'\confidence.txt', 'a') as f:f.write(f'{p.stem} ' + ('%g ' * len(line)).rstrip() % line + f' {conf}\n')# Print time (inference + NMS)print(f'{s}')print(f'这张图片搞定了！inference 与NMS共耗时：{t2 - t1:.3f}秒\n')# Stream results:是否需要显示预测后的结果  img0(此时已将pred结果可视化到了img0中)# TODO --view-imgif view_img:if names[int(cls)] in Aclasses:cv2.imshow(str(p), im0)cv2.waitKey(2)  # 图片显示停留时间# Save results (image with detections):是否需要保存图片或视频（检测后的图片/视频 里面已经被画好了框的） img0if save_img:if names[int(cls)] in Aclasses:cv2.imwrite(save_path, im0)# 保存原图save_im0s = 1if save_im0s:cv2.imwrite(save_path, im0s)else:  # 'video' or 'stream'if names[int(cls)] in Aclasses:if vid_path != save_path:  # new videovid_path = save_pathif isinstance(vid_writer, cv2.VideoWriter):vid_writer.release()  # release previous video writerif vid_cap:  # videofps = vid_cap.get(cv2.CAP_PROP_FPS)w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))else:  # streamfps, w, h = 30, im0.shape[1], im0.shape[0]save_path += '.mp4'vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h))vid_writer.write(im0)if save_txt or save_img:s = f"\n{len(list(save_dir.glob('labels/*.txt')))} 个标签已经保存到了以下地址： {save_dir / 'labels'}" if save_txt else ''print(f"{s}")print(f'检测程序结束运行，总共耗时：{time.time() - t0:.3f}秒')return expertPath# parameters configurationparser = argparse.ArgumentParser()# 模型参数（pt文件）地址parser.add_argument('--weights', nargs='+', type=str, default=Model_path, help='model.pt path(s)')# 待检测目标（会读取文件夹下所有）parser.add_argument('--source', type=str, default=Input_path, help='source')  # file/folder, 0 for webcam# 输入图像大小，最好和网络模型对应（640或1280）parser.add_argument('--img-size', type=int, default=Img_size, help='inference size (pixels)')# 检测框的置信度（有多大把握把人物正好框进去）parser.add_argument('--conf-thres', type=float, default=Confidence, help='object confidence threshold')# 非极大抑制（IOU代表检测框的重合度）parser.add_argument('--iou-thres', type=float, default=IOU, help='IOU threshold for NMS')# 硬件选择（CPU还是GPU）parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')# 以下需可以在  编辑配置  中添加  形参  以决定False or True（默认为False，输入--view-img这一项便为True：view-img = True）# 是否显示检测过程（img by img）parser.add_argument('--view-img', action='store_true', help='display results')# 是否在保存labels文件.txtparser.add_argument('--save-txt', action='store_true', help='save results to *.txt')# 是否保存置信度parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels')# 是否保存被检测对象parser.add_argument('--nosave', action='store_true', help='do not save images/videos')# 是否增强NMS，会更加强大、提升结果parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS')# 同是增强检测效果parser.add_argument('--augment', action='store_true', help='augmented inference')# 为True只保留本模型需要的东西parser.add_argument('--update', action='store_true', help='update all Models')# 保存文件同名时是否新增一个文件夹保存（为True会保存在同名的文件夹下）parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment')# 决定保留class多少的结果（--class 0）parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --class 0, or --class 0 2 3')# 保存地址parser.add_argument('--project', default=Save_path, help='save results to project/name')# 保存文件名parser.add_argument('--name', default=Save_name, help='save results to project/name')opt = parser.parse_args()print(f'parser.parse_args(分析程序的参数):\n{opt}')# check_requirements(exclude=('pycocotools', 'thop'))with torch.no_grad():if opt.update:  # 更新所有模型(修复SourceChangeWarning)for opt.weights in ['yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt']:ExpertPath = detect()strip_optimizer(opt.weights)else:ExpertPath = detect()print(f'\nHi,congratulation!Don t forget 'f'\n1.pt file path:{Model_path}'f'\n2.imgs path:{Input_path}'f'\nNow,you can see the result in {Save_path}/{Save_name}'f'\nAnd then you can fix these labels in labelling or webs such as makesence! 'f'Because human brain is the best computer for now,lol ^_^')if input(f'\nDo you want to save original photos with their labels in save directory?[y/n]:') == 'y':# TODO: save original photos with their labelsSave_path = Save_path + '/' + Save_nameprint(f'以下是直接参与后续微调的图片:')save_oriImgs(Save_path, Input_path)print(f'以下是expertLabel images:')save_oriImgs(ExpertPath, Input_path)else:print(f'\nHi hi hi, therefore you don t want to save original  photos, I have to remind you again: \nORIGINAL PHOTOS PATH:{Input_path}')if input(f'\nAre you sure to reconfiguration[y/n]:') == 'y':breakreturn Save_pathif __name__ == '__main__':#global ExpertPath# cuda可用性检查CudaIsAvailable()# main recurrentwhile True:datasetPath = search()print('\nIf you don t want to exit please [input n], or you can reconfiguration:')if input(f'Are you sure to exit:') == 'exit':break# 打开标注网站（以便人工修改）Relabelled = input(f'\nHi hi,I forget one thing ! Do you want to Manual finishing labeling now?[y/n]:')if Relabelled == 'y':# labelling Web introduction blog: https://blog.csdn.net/To_ChaRiver/article/details/119619515OpenLabelWeb()print(f'See you next time and dont forget you dataset are sleeping in :{datasetPath} !')

依赖版本：Yolov5 5.0

ps. class类别选择


Aclasses = ['background']
while True:# 想要检测的类别while True:while True:transNum = input(f'number or name[1/2]：')if transNum == '1':Class_name = int(input('input the class you want to detect:'))breakelif transNum == '2':Class_name = input('input the class you want to detect:')breakc = Class_nameprint(f'Class_name:{c}')bool1 = c in class_numbersprint(f'bool1:{bool1}')bool2 = c in class_names  # 不一定要bool，bl也行（任何字符串转一下就行）print(f"bool2:{bool2}")# if bool == True:  # Class_name in class_numbers == True一直返回False就很奇怪bool = bool1 or bool2print(f'输入判别：{bool}')if bool1:print(f'即将检测含有{c}的图片>_<\n')breakelse:if bool2:print(f'即将检测含有{class_names[c]}的图片')breakelse:print('\n请输入class中有的目标名称或编号')if transNum == '1':class_input = class_names[int(c)]if class_input not in Aclasses:Aclasses.append(class_input)else:if c not in Aclasses:Aclasses.append(c)enough = input('enough [input y] ?')if enough == 'y':breakprint(f'{Aclasses},so go on')print(f'classes that you choose:{Aclasses}')

Yolov5官方网络改进：增加search模块（基于迁移学习的目标检测+多模态零样本自定义标签分类网络）相关推荐

MXnet-gluoncv实现基于迁移学习的目标检测（自己的数据集）
一.安装首先安装mxnet(cpu版): 命令行输入pip install -i https://pypi.doubanio.com/simple/ mxnet(豆瓣源.速度快) 接着安装gluon ...
基于深度学习的目标检测综述（一）：简介及骨干网络
这篇文献的主要结构如下: 第1节,简单介绍目标检测的发展及两类目标检测算法. 第 2 节中讨论了骨干网络,目标检测器需要强大的骨干网络来提取丰富的特征.众所周知,特定领域图像检测器的典型管道是任务的基 ...
病虫害模型算法_基于深度学习的目标检测算法综述
sigai 基于深度学习的目标检测算法综述导言目标检测的任务是找出图像中所有感兴趣的目标(物体),确定它们的位置和大小,是机器视觉领域的核心问题之一.由于各类物体有不同的外观,形状,姿态,加上成像 ...
基于深度学习的目标检测的研究进展2
普通的深度学习监督算法主要是用来做分类,如图1(1)所示,分类的目标是要识别出图中所示是一只猫.而在ILSVRC(ImageNet Large Scale Visual Recognition Cha ...
学习笔记之——基于深度学习的目标检测算法
国庆假期闲来无事~又正好打算入门基于深度学习的视觉检测领域,就利用这个时间来写一份学习的博文~本博文主要是本人的学习笔记与调研报告(不涉及商业用途),博文的部分来自我团队的几位成员的调研报告(由于隐私 ...
基于深度学习的目标检测综述(单阶段、多阶段、FPN变体、旋转目标检测等)
随着深度学习的发展,基于深度学习的目标检测方法因其优异的性能已经得到广泛的使用.目前经典的目标检测方法主要包括单阶段(YOLO.SSD.RetinaNet,还有基于关键点的检测方法等)和多阶段方法(F ...
基于深度学习的目标检测研究综述
基于深度学习的目标检测研究综述摘要:深度学习是机器学习的一个研究领域,近年来受到越来越多的关注.最近几年,深度学习在目标检测领域取得了不少突破性的进展,已经运用到具体的目标检测任务上.本文首先详细介 ...
基于深度学习的目标检测算法综述(从R-CNN到Mask R-CNN)
深度学习目标检测模型全面综述:Faster R-CNN.R-FCN和SSD 从RCNN到SSD,这应该是最全的一份目标检测算法盘点基于深度学习的目标检测算法综述(一) 基于深度学习的目标检测算法综述 ...
基于深度学习的目标检测算法综述（一）
基于深度学习的目标检测算法综述(一) 基于深度学习的目标检测算法综述(二) 基于深度学习的目标检测算法综述(三) 本文内容原创,作者:美图云视觉技术部检测团队,转载请注明出处目标检测(Object ...

Yolov5官方网络改进：增加search模块（基于迁移学习的目标检测+多模态零样本自定义标签分类网络）

Yolov5官方网络改进：增加search模块（基于迁移学习的目标检测+多模态零样本自定义标签分类网络）相关推荐

最新文章

热门文章