使用YOLOv4进行自定义数据集训练的流程及结果_yolov4训练过程

使用YOLOv4进行自定义数据集训练的流程及结果

一、安装与配置

• git clone Darknet github

git clone https://github.com/AlexeyAB/darknet
1

• 修改Makefile

​ 进入 darknet_master 文件架修改 Makefile，将GPU, CUDNN, CUDNN_HALF, OPENCV 修改为1，预设值为0

• 进行编译

cd darknet
make
1
2

二、项目创建

• 先将数据集图片暂存至临时文件夹中

• 创建如图的文件结构

  

• 使用labelimg给图片数据集进行标注

  1. 将/labelimg/data/predefined_classes.txt中修改为自定义数据集中的类
  2. 运行labelimg进行打标，AD左右切换图片，W创建标注框，将标注的VOC文件存储到项目临时文件夹中

• 使用数据集增强功能包进行数据集增强

  1. 将数据集的标注和标注好的标注文件分别存储到功能包DataAugForObjectDetecton中的/data/Annotations和/data/images中
  2. 进入DataAugForObjectDetecton.py文件，到程序入口处修改need_aug_num = 10,这是每张图片进行扩增的数量
  3. 运行DataAugForObjectDetecton.py文件，在Dataset文件夹下可以找到扩增后的标签与数据集
  4. 将标签与数据集移动到项目中对应的文件夹

• 划分训练集、测试集和验证集

  运行spit.py文件，调节其中的参数可以改变三者的比例大小

• VOC标签转YOLO

  运行voc_lavel.py文件，转化后的标签将存储在label文件夹中

• 写入2007_test.txt(2007_train.txt 2007_val.txt)

  运行write_name.py文件，可以进行文件的写入(读取spit.py生成的train.txt test.txt val.txt中的内容并且将补全的路径输出到指定的文件中)

# write_name.py
import os

def write_name(name_file_path,path_file_path,task_path):

    name_file = open(name_file_path , 'r' , encoding= 'utf-8')
    path_file = open(path_file_path , 'w' , encoding= 'utf-8')
    for i in name_file:
        i = i[:-1]
        path = f"{task_path}/{i}.jpg\n"
        path_file.write(path)
    path_file.close()
    name_file.close()

if __name__ == "_main__":
    task_path = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/VOC2007/JPEGImages"
    name_file_path1 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/VOC2007/ImageSets/Main/train.txt"
    path_file_path1 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_train.txt"
    name_file_path2 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/VOC2007/ImageSets/Main/test.txt"
    path_file_path2 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_test.txt"
    name_file_path3 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/VOC2007/ImageSets/Main/val.txt"
    path_file_path3 = "/home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_val.txt"    
    write_name(name_file_path1,path_file_path1,task_path)
    write_name(name_file_path2,path_file_path2,task_path)
    write_name(name_file_path3,path_file_path3,task_path)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

• 修改yolov4-tiny.cfg文件

  classes #改为自定义的类别数
  filters = (classes + 5)*3
  1
  2

• voc.names文件
  只需要将类别分行写在其中，以我的项目为例：

# voc.names文件
corn
corn_plant
cucumber
cucumber_plant
watermelon
rice
wheat
1
2
3
4
5
6
7
8

• voc.data文件（以我的项目为例，每一行都需要对应的修改）

classes = 7 #类别数
train  = /home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_train.txt	#存放train图片路径的txt
valid  = /home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_val.txt
test   = /home/ding/deeplearning_task/winter_task5/VOCdevkit/2007_test.txt
names = /home/ding/deeplearning_task/winter_task5/VOCdevkit/voc.names		#voc.names路径
backup = /home/ding/deeplearning_task/winter_task5/VOCdevkit/VOC2007/backup	#训练权重放置的路径
1
2
3
4
5
6

三、模型训练与验证操作

• 相关操作指令

  ./darknet detector train [.data] [.cfg] [.weight]
  # 模型训练
  1
  2

  说明：
  .data是.data文件的路径。

  .cfg是.cfg文件的路径。

  .weight是预训练权重文件的路径，可以是.weight文件，也可以是.backup文件。

  ./darknet detector train [.data] [.cfg] [.weight]
  #检测图片
  ./darknet detector demo [.data] [.cfg] [.weight] [-thresh 0.25] [xxx.mp4]
  # 检测视频
  -dont_show -ext_output [.txt] result.txt
  #批量检测
  ./darknet detector demo [.data] [.cfg] [.weight] [-thresh 0.25] [http://192.168.1.108:8080/video?dummy=x.mjpg -i 0]
  #检测网络摄像头
  ./darknet detector map [.data] [.cfg] [.weight]
  #评价模型
  1
  2
  3
  4
  5
  6
  7
  8
  9
  10

四、任务模型训练结果

• 类别划分

我的训练划分了七个类别

corn
corn_plant
cucumber
cucumber_plant
watermelon
rice
wheat
1
2
3
4
5
6
7

• 训练过程损失曲线

  

• 图片验证结果