最近在学图像识别，跑了几遍yolov3，在此做一些记录

我的环境如下：

ubuntu: 18.04
GPU: RTX3050ti
cuda: 11.4
cudnn: 8.4.1
opencv: 4.2
python: 3.6.9

首先是框架的安装，我选择的是AB大神的darknet框架，下载过程如下：

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

修改一下makefile文件

GPU=1//GPU=1表示启用GPU
CUDNN=1//CUDNN=1表示启用cudnn
CUDNN_HALF=0
OPENCV=1//opencv=1表示启用opencv，如果需要调用摄像头需启用
AVX=0
OPENMP=0
LIBSO=0
ZED_CAMERA=0
ZED_CAMERA_v2_8=0

# set GPU=1 and CUDNN=1 to speedup on GPU
# set CUDNN_HALF=1 to further speedup 3 x times (Mixed-precision on Tensor Cores) GPU: Volta, Xavier, Turing and higher
# set AVX=1 and OPENMP=1 to speedup on CPU (if error occurs then set AVX=0)
# set ZED_CAMERA=1 to enable ZED SDK 3.0 and above
# set ZED_CAMERA_v2_8=1 to enable ZED SDK 2.X

我的建议是一定要用cuda，纯CPU跑的特别慢，可以考虑云GPU

编译：

make

下载yolov3预训练模型：（如果下载的很慢的话可以从博主的这篇博客下载）

wget https://pjreddie.com/media/files/yolov3.weights

测试一下是否安装好了：

./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg

或者

./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg

运行结果如下时，恭喜你安装成功了

接下来就可以着手训练自己的数据集了

一.训练数据集的第一步首先是标注图像，我这里选用的是labelImg,下载地址，安装成功后终端输入：(建议使用conda虚拟环境）

labelImg

打开标注界面如下：

yolo使用的文本格式是txt文件，但你也可以标注为xml文件，之后再进行转换即可

二.接下来在darknet目录下创建myData文件夹，目录结构如下，将之前标注好的图片和xml文件放到对应目录下

myData
  ...JPEGImages#存放图像
  
  ...ImageSets/Main #(如果标注的是xml文件的话需创建，txt文件不需要）
 
  ...Annotations#存放图像对应的xml文件（如果标注的是xml文件的话）
 
  ...labels #存放图像对应的txt文件

  ...backup #存放训练所得的权重文件

三.数据集的配置

     1.如果标注的是xml文件，新建my_labels.py文件，复制以下内容：这里参考

import os
import random
 
trainval_percent = 0.1
train_percent = 0.9
#xmlfilepath = 'Annotations'
filepath = 'labels'
txtsavepath = 'ImageSets\Main'
#total_xml = os.listdir(xmlfilepath)
total = os.listdir(filepath)
#num = len(total_xml)
num = len(total)
list = range(num)
tv = int(num * trainval_percent)
tr = int(tv * train_percent)
trainval = random.sample(list, tv)
train = random.sample(trainval, tr)
 
ftrainval = open('ImageSets/Main/trainval.txt', 'w')
ftest = open('ImageSets/Main/test.txt', 'w')
ftrain = open('ImageSets/Main/train.txt', 'w')
fval = open('ImageSets/Main/val.txt', 'w')
 
for i in list:
    #name = total_xml[i][:-4] + '\n'
    name = total[i][:-4] + '\n'
    if i in trainval:
        ftrainval.write(name)
        if i in train:
            ftest.write(name)
        else:
            fval.write(name)
    else:
        ftrain.write(name)
 
ftrainval.close()
ftrain.close()
fval.close()
ftest.close()

         运行：

python my_labels.py

       会生成labels文件夹，里面放的是对应的txt文件 ;同时在 ImageSets/Main路径下会生成train.txt和test.txt文件

   2.如果标注的是txt格式，新建labels.py文件,复制以下内容：

# -*- coding: utf-8 -*-
# 此代码和myData文件夹同目录
import glob
import os, random, shutil
import sys, getopt
import string


#    train, test, rate = getDir(sys.argv[1:])
# if tmp <= 0.0 or tmp >= 1.0:
#   rate = 0.1

def moveFile(trainDir, testDir, rate):
    rate=float(rate)
    pathDir = os.listdir(trainDir)          #返回指定的文件夹包含的文件或文件夹的名字的列表
    filenumber=len(pathDir)
    print("filenumber = ", filenumber)
    picknumber=int(filenumber*rate)
    print("picknumber = ", picknumber)
    sample = random.sample(pathDir, picknumber)          #从pathDir中随机选取picknumber个元素
    for name in sample:
        shutil.move(os.path.join(trainDir,name), os.path.join(testDir,name))
    return

train_list = []
test_list = []
train_file = 'train.txt'
test_file = 'test.txt'
rate = 0.80

if __name__ == '__main__':

    rate = float(rate)
    #pathDir = os.listdir('labels/')
    pathDir = os.listdir('/home/your/darknet/myData/JPEGImages/')#改为你自己的路径
    filenumber = len(pathDir)
    picknumber = int(filenumber * rate)

    sample = random.sample(pathDir, picknumber)
    for name in sample:
        train_list.append(name)

    for name in pathDir:
        if name not in sample:
            test_list.append(name)

    cur_dir = os.getcwd()                #返回当前进程的工作目录
    train_images_dir = os.path.join(cur_dir, 'JPEGImages/')

    with open(train_file, 'w') as train_txt:
        for name in train_list:
            jpg_name = name.strip()
            jpg_file = os.path.join(train_images_dir, jpg_name)
            train_txt.write(jpg_file + '\n')
    train_txt.close()

    with open(test_file, 'w') as test_txt:
        for name in test_list:
            jpg_name = name.strip()
            jpg_file = os.path.join(train_images_dir, jpg_name)
            test_txt.write(jpg_file + '\n')
    test_txt.close()

    print(filenumber,picknumber,filenumber-picknumber)

     运行：

python labels.py

    会在myData文件夹下生成test.txt和train.txt文件

3.在myData文件夹下新建myData.names文件，内容如下：

#根据自己的数据集标注的标签，按照序号顺序填写
tissue
roll-of-paper
battled-drinks
chewing-gum
banana

四.cfg文件和data文件的修改

  在cfg文件夹中创建my_data.data文件，复制以下内容：

classes= 5 ##改为自己的分类个数
##下面都改为自己的路径
train  = /home/your/darknet/myData/train.txt #你的train.txt和test.txt在哪里就改成相应路径
valid  = /home/your/darknet/myData/test.txt
names = /home/your/darknet/myData/myData.names
backup = /home/your/darknet/myData/backup/

  复制cfg文件夹中的yolov3.cfg

  以下是cfg文件参数详解，参考

  我的显卡只能支持batch=64时subdvision=64，如果出现cuda out of memory错误，可以减小batch或增大subdvision ；max_batches根据个人需求修改，如果数据集不大的话可以设置小一些，我的数据集100-500张图片的基本都在1200左右达到不错的效果，1500-1800左右收敛，所以当数据集不大时可以考虑修改为2000，这是我自己的拙见。

[net]
# Testing                                  测试模式
# batch=1
# subdivisions=1
# Training                                 训练模式
 batch=64                                  一批训练样本的样本数量，每batch个样本更新一次参数
 subdivisions=16                           batch/subdivisions作为一次性送入训练器的样本数量,如果内存不够大，将batch分割为subdivisions个子batch
 
										   上面这两个参数如果电脑内存小，则把batch改小一点，batch越大，训练效果越好
										   subdivisions越大，可以减轻显卡压力
										   
										   
width=416                                  input图像的宽
height=416                                 input图像的高
channels=3                                 input图像的通道数
 
                                           以上三个参数为输入图像的参数信息 width和height影响网络对输入图像的分辨率，
					                       从而影响precision，只可以设置成32的倍数
										   
										   
momentum=0.9                               [?]DeepLearning1中最优化方法中的动量参数，这个值影响着梯度下降到最优值得速度https://nanfei.ink/2018/01/23/YOLOv2%E8%B0%83%E5%8F%82%E6%80%BB%E7%BB%93/#more
 
 
decay=0.0005                               [?]权重衰减正则项，防止过拟合.每一次学习的过程中，将学习后的参数按照固定比例进行降低，为了防止过拟合，decay参数越大对过拟合的抑制能力越强。
 
 
angle=0                                    通过旋转角度来生成更多训练样本
saturation = 1.5                           通过调整饱和度来生成更多训练样本
exposure = 1.5                             通过调整曝光量来生成更多训练样本
hue=.1                                     通过调整色调来生成更多训练样本
 
 
 
learning_rate=0.001                        学习率决定着权值更新的速度，设置得太大会使结果超过最优值，太小会使下降速度过慢。
										   如果仅靠人为干预调整参数，需要不断修改学习率。刚开始训练时可以将学习率设置的高一点，
										   而一定轮数之后，将其减小
										   在训练过程中，一般根据训练轮数设置动态变化的学习率。
										   刚开始训练时：学习率以 0.01 ~ 0.001 为宜。
										   一定轮数过后：逐渐减缓。
										   接近训练结束：学习速率的衰减应该在100倍以上。
										   学习率的调整参考https://blog.csdn.net/qq_33485434/article/details/80452941
										   
										   
burn_in=1000                               在迭代次数小于burn_in时，其学习率的更新有一种方式，大于burn_in时，才采用policy的更新方式
max_batches = 20200                        训练达到max_batches后停止学习
policy=steps                               这个是学习率调整的策略，有policy：constant, steps, exp, poly, step, sig, RANDOM，constant等方式
					                       参考https://nanfei.ink/2018/01/23/YOLOv2%E8%B0%83%E5%8F%82%E6%80%BB%E7%BB%93/#more
										   
										   
steps=40000,45000                          下面这两个参数steps和scale是设置学习率的变化，比如迭代到40000次时，学习率衰减十倍。
scales=.1,.1                               45000次迭代时，学习率又会在前一个学习率的基础上衰减十倍
 
 
 
[convolutional]
batch_normalize=1                          是否做BN
filters=32                                 输出特征图的数量
size=3                                     卷积核的尺寸
stride=1                                   做卷积运算的步长
pad=1                                      如果pad为0,padding由 padding参数指定;如果pad为1，padding大小为size/2，padding应该是对输入图像左边缘拓展的像素数量
activation=leaky                           激活函数的类型
 
# Downsample
 
[convolutional]
batch_normalize=1
filters=64
size=3
stride=2
pad=1
activation=leaky
 
[convolutional]
batch_normalize=1
filters=32
size=1
stride=1
pad=1
activation=leaky
 
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
 
[shortcut]
from=-3
activation=linear
 
# Downsample
 
   ......
 
# Downsample
 
 
######################
 
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
 
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
 
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
 
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
 
[convolutional]
batch_normalize=1
filters=512
size=1
stride=1
pad=1
activation=leaky
 
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
 
[convolutional]
size=1
stride=1
pad=1
filters=45                          每一个[region/yolo]层前的最后一个卷积层中的 filters=(classes+1+coords)*anchors_num，
                                    其中anchors_num 是该层mask的一个值.如果没有mask则 anchors_num=num是这层的ancho
					                5的意义是5个坐标，论文中的tx,ty,tw,th,to
activation=linear
 
[yolo]                              在yoloV2中yolo层叫region层
mask = 6,7,8                        这一层预测第6、7、8个 anchor boxes ，每个yolo层实际上只预测3个由mask定义的anchors
 
 
anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326
                                    [?]anchors是可以事先通过cmd指令计算出来的，是和图片数量，width,height以及cluster(应该就是下面的num的值，
					                即想要使用的anchors的数量)相关的预选框，可以手工挑选，也可以通过kmeans 从训练样本中学出
									
									
classes=10                          网络需要识别的物体种类数
num=9                               每个grid cell预测几个box,和anchors的数量一致。当想要使用更多anchors时需要调大num，且如果调大num后训练时Obj趋近0的话可以尝试调大object_scale
jitter=.3                           [?]利用数据抖动产生更多数据，YOLOv2中使用的是crop，filp，以及net层的angle，flip是随机的，
					                jitter就是crop的参数，tiny-yolo-voc.cfg中jitter=.3，就是在0~0.3中进行crop
									
									
ignore_thresh = .5                  决定是否需要计算IOU误差的参数，大于thresh，IOU误差不会夹在cost function中 
truth_thresh = 1
random=0                            如果为1，每次迭代图片大小随机从320到608，步长为32，如果为0，每次训练大小与输入大小一致
 
[route]
layers = -4
 
......
 
 
#可以添加没有标注框的图片和其空的txt文件，作为negative数据
#可以在第一个[yolo]层之前的倒数第二个[convolutional]层末尾添加 stopbackward=1,以此提升训练速度
#即使在用416*416训练完之后，也可以在cfg文件中设置较大的width和height，增加网络对图像的分辨率，从而更可能检测出图像中的小目标，而不需要重新训练
#Out of memory的错误需要通过增大subdivisions来解决

修改yolo层和yolo层上的convolutional层的参数，具体为：

convolutional层中：
filters=(classes+1+coords)*anchors_num  一般为filters=(classes+5)*3
yolo层中
classes=数据集中的物体个数
#一共要改3处

可以改变yolo层中的anchor的值使结果精度更高，但改不改好像没什么区别，如果要修改的话，在myData文件夹下新建anchors.py文件，复制以下内容：参考

# -*- coding: utf-8 -*-
import numpy as np
import random
import argparse
import os
#参数名称
parser = argparse.ArgumentParser(description='使用该脚本生成YOLO-V3的anchor boxes\n')
parser.add_argument('--input_annotation_txt_dir',required=True,type=str,help='输入存储图片的标注txt文件(注意不要有中文)')
parser.add_argument('--output_anchors_txt',required=True,type=str,help='输出的存储Anchor boxes的文本文件')
parser.add_argument('--input_num_anchors',required=True,default=6,type=int,help='输入要计算的聚类（Anchor boxes的个数）')
parser.add_argument('--input_cfg_width',required=True,type=int,help="配置文件中width")
parser.add_argument('--input_cfg_height',required=True,type=int,help="配置文件中height")
args = parser.parse_args()
'''
centroids 聚类点 尺寸是 numx2,类型是ndarray
annotation_array 其中之一的标注框
'''
def IOU(annotation_array,centroids):
    #
    similarities = []
    #其中一个标注框
    w,h = annotation_array
    for centroid in centroids:
        c_w,c_h = centroid
        if c_w >=w and c_h >= h:#第1中情况
            similarity = w*h/(c_w*c_h)
        elif c_w >= w and c_h <= h:#第2中情况
            similarity = w*c_h/(w*h + (c_w - w)*c_h)
        elif c_w <= w and c_h >= h:#第3种情况
            similarity = c_w*h/(w*h +(c_h - h)*c_w)
        else:#第3种情况
            similarity = (c_w*c_h)/(w*h)
        similarities.append(similarity)
    #将列表转换为ndarray
    return np.array(similarities,np.float32) #返回的是一维数组，尺寸为(num,)
 
'''
k_means:k均值聚类
annotations_array 所有的标注框的宽高，N个标注框，尺寸是Nx2,类型是ndarray
centroids 聚类点 尺寸是 numx2,类型是ndarray
'''
def k_means(annotations_array,centroids,eps=0.00005,iterations=200000):
    #
    N = annotations_array.shape[0]#C=2
    num = centroids.shape[0]
    #损失函数
    distance_sum_pre = -1
    assignments_pre = -1*np.ones(N,dtype=np.int64)
    #
    iteration = 0
    #循环处理
    while(True):
        #
        iteration += 1
        #
        distances = []
        #循环计算每一个标注框与所有的聚类点的距离（IOU）
        for i in range(N):
            distance = 1 - IOU(annotations_array[i],centroids)
            distances.append(distance)
        #列表转换成ndarray
        distances_array = np.array(distances,np.float32)#该ndarray的尺寸为 Nxnum
        #找出每一个标注框到当前聚类点最近的点
        assignments = np.argmin(distances_array,axis=1)#计算每一行的最小值的位置索引
        #计算距离的总和，相当于k均值聚类的损失函数
        distances_sum = np.sum(distances_array)
        #计算新的聚类点
        centroid_sums = np.zeros(centroids.shape,np.float32)
        for i in range(N):
            centroid_sums[assignments[i]] += annotations_array[i]#计算属于每一聚类类别的和
        for j in range(num):
            centroids[j] = centroid_sums[j]/(np.sum(assignments==j))
        #前后两次的距离变化
        diff = abs(distances_sum-distance_sum_pre)
        #打印结果
        print("iteration: {},distance: {}, diff: {}, avg_IOU: {}\n".format(iteration,distances_sum,diff,np.sum(1-distances_array)/(N*num)))
        #三种情况跳出while循环：1：循环20000次，2：eps计算平均的距离很小 3：以上的情况
        if (assignments==assignments_pre).all():
            print("按照前后两次的得到的聚类结果是否相同结束循环\n")
            break
        if diff < eps:
            print("按照eps结束循环\n")
            break
        if iteration > iterations:
            print("按照迭代次数结束循环\n")
            break
        #记录上一次迭代
        distance_sum_pre = distances_sum
        assignments_pre = assignments.copy()
if __name__=='__main__':
    #聚类点的个数，anchor boxes的个数
    num_clusters = args.input_num_anchors
    #索引出文件夹中的每一个标注文件的名字(.txt)
    names = os.listdir(args.input_annotation_txt_dir)
    #标注的框的宽和高
    annotations_w_h = []
    for name in names:
        txt_path = os.path.join(args.input_annotation_txt_dir,name)
        #读取txt文件中的每一行
        f = open(txt_path,'r')
        for line in f.readlines():
            line = line.rstrip('\n')
            w,h = line.split(' ')[3:]#这时读到的w,h是字符串类型
            #eval()函数用来将字符串转换为数值型
            annotations_w_h.append((eval(w),eval(h)))
        f.close()
        #将列表annotations_w_h转换为numpy中的array,尺寸是(N,2),N代表多少框
        annotations_array = np.array(annotations_w_h,dtype=np.float32)
    N = annotations_array.shape[0]
    #对于k-means聚类，随机初始化聚类点
    random_indices = [random.randrange(N) for i in range(num_clusters)]#产生随机数
    centroids = annotations_array[random_indices]
    #k-means聚类
    k_means(annotations_array,centroids,0.00005,200000)
    #对centroids按照宽排序，并写入文件
    widths = centroids[:,0]
    sorted_indices = np.argsort(widths)
    anchors = centroids[sorted_indices]
    #将anchor写入文件并保存
    f_anchors = open(args.output_anchors_txt,'w')
    #
    for anchor in  anchors:
        f_anchors.write('%d,%d'%(int(anchor[0]*args.input_cfg_width),int(anchor[1]*args.input_cfg_height)))
        f_anchors.write('\n')

五.开始训练

1.下载预训练文件（有的话更好，没有也能跑）终端输入： （如果下载的很慢的话可以从博主的这篇博客下载）

wget https://pjreddie.com/media/files/darknet53.conv.74

2.开始训练，终端输入：

./darknet detector train cfg/my_data.data cfg/my_yolov3.cfg darknet53.conv.74

或者指定gpu训练，默认使用gpu0

./darknet detector train cfg/my_data.data cfg/my_yolov3.cfg darknet53.conv.74 -gups 0,1,2,3

 3.等待训练完成，在myData文件夹下的backup文件夹(需要自己创建）中寻找这一炉仙丹吧

 4.如果中途因为out of memory中断进程，修改batches和subdvision 重新训练，从中断处开始训练：

./darknet detector train cfg/my_data.data cfg/my_yolov3.cfg myData/backup/my_yolov3.backup -gpus 0,1,2,3

我的效果图（1200的时候有点事退出去了，最后效果还可以，但avg loss达到0.06左右效果更好：

六.测试

终端输入：

#图片测试，视频检测同理
./darknet detector test cfg/my_data.data cfg/my_yolov3.cfg myData/backup/my_yolov3_last.weights 1.jpg
#1.jpg改为图片路径+图片名，我直接在darknet文件夹下检测的，所以没有加路径
#或者运行darknet文件夹下的darknet_images.py文件，记得修改路径
python3 darknet_images.py
#摄像头测试
./darknet detector demo cfg/my_data.data cfg/my_yolov3.cfg myData/backup/my_yolov3_last.weights 

我的最终效果:

七.批量测试

在darknet文件夹下创建detect.py,复制以下内容（注意修改为自己的路径）：

import argparse
import os
import glob
import random
import darknet
import time
import cv2
import numpy as np
import darknet



def parser():
    parser = argparse.ArgumentParser(description="YOLO Object Detection")
    parser.add_argument("--input", type=str, default="",
                        help="image source. It can be a single image, a"
                        "txt with paths to them, or a folder. Image valid"
                        " formats are jpg, jpeg or png."
                        "If no input is given, ")
    parser.add_argument("--batch_size", default=1, type=int,
                        help="number of images to be processed at the same time")
    parser.add_argument("--weights", default="myData/backup/my_yolov3_last.weights",#修改为自己的路径
                        help="yolo weights path")
    parser.add_argument("--dont_show", action='store_true',
                        help="windown inference display. For headless systems")
    parser.add_argument("--ext_output", action='store_true',
                        help="display bbox coordinates of detected objects")
    parser.add_argument("--save_labels", action='store_true',
                        help="save detections bbox for each image in yolo format")
    parser.add_argument("--config_file", default="./cfg/my_yolov3.cfg",
                        help="path to config file")
    parser.add_argument("--data_file", default="./cfg/my_data.data",
                        help="path to data file")
    parser.add_argument("--thresh", type=float, default=.25,
                        help="remove detections with lower confidence")
    return parser.parse_args()

def check_arguments_errors(args):
    assert 0 < args.thresh < 1, "Threshold should be a float between zero and one (non-inclusive)"
    if not os.path.exists(args.config_file):
        raise(ValueError("Invalid config path {}".format(os.path.abspath(args.config_file))))
    if not os.path.exists(args.weights):
        raise(ValueError("Invalid weight path {}".format(os.path.abspath(args.weights))))
    if not os.path.exists(args.data_file):
        raise(ValueError("Invalid data file path {}".format(os.path.abspath(args.data_file))))
    if args.input and not os.path.exists(args.input):
        raise(ValueError("Invalid image path {}".format(os.path.abspath(args.input))))


def check_batch_shape(images, batch_size):
    """
        Image sizes should be the same width and height
    """
    shapes = [image.shape for image in images]
    if len(set(shapes)) > 1:
        raise ValueError("Images don't have same shape")
    if len(shapes) > batch_size:
        raise ValueError("Batch size higher than number of images")
    return shapes[0]


def load_images(images_path):
    """
    If image path is given, return it directly
    For txt file, read it and return each line as image path
    In other case, it's a folder, return a list with names of each
    jpg, jpeg and png file
    """
    input_path_extension = images_path.split('.')[-1]
    if input_path_extension in ['jpg', 'jpeg', 'png']:
        return [images_path]
    elif input_path_extension == "txt":
        with open(images_path, "r") as f:
            return f.read().splitlines()
    else:
        return glob.glob(
            os.path.join(images_path, "*.jpg")) + \
            glob.glob(os.path.join(images_path, "*.png")) + \
            glob.glob(os.path.join(images_path, "*.jpeg"))


def prepare_batch(images, network, channels=3):
    width = darknet.network_width(network)
    height = darknet.network_height(network)

    darknet_images = []
    for image in images:
        image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
        image_resized = cv2.resize(image_rgb, (width, height),
                                   interpolation=cv2.INTER_LINEAR)
        custom_image = image_resized.transpose(2, 0, 1)
        darknet_images.append(custom_image)

    batch_array = np.concatenate(darknet_images, axis=0)
    batch_array = np.ascontiguousarray(batch_array.flat, dtype=np.float32)/255.0
    darknet_images = batch_array.ctypes.data_as(darknet.POINTER(darknet.c_float))
    return darknet.IMAGE(width, height, channels, darknet_images)


def image_detection(image_path,network, class_names, class_colors, thresh):
    # Darknet doesn't accept numpy images.
    # Create one with image we reuse for each detect
    width = darknet.network_width(network)
    height = darknet.network_height(network)
    darknet_image = darknet.make_image(width, height, 3)
   
    image = cv2.imread(image_path)
    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image_resized = cv2.resize(image_rgb, (width, height),
                               interpolation=cv2.INTER_LINEAR)

    darknet.copy_image_from_bytes(darknet_image, image_resized.tobytes())
    detections = darknet.detect_image(network, class_names, darknet_image, thresh=thresh)
    darknet.free_image(darknet_image)
    image = darknet.draw_boxes(detections, image_resized, class_colors)
    return cv2.cvtColor(image, cv2.COLOR_BGR2RGB), detections


def batch_detection(network, images, class_names, class_colors,
                    thresh=0.25, hier_thresh=.5, nms=.45, batch_size=4):
    image_height, image_width, _ = check_batch_shape(images, batch_size)
    darknet_images = prepare_batch(images, network)
    batch_detections = darknet.network_predict_batch(network, darknet_images, batch_size, image_width,
                                                     image_height, thresh, hier_thresh, None, 0, 0)
    batch_predictions = []
    for idx in range(batch_size):
        num = batch_detections[idx].num
        detections = batch_detections[idx].dets
        if nms:
            darknet.do_nms_obj(detections, num, len(class_names), nms)
        predictions = darknet.remove_negatives(detections, class_names, num)
        images[idx] = darknet.draw_boxes(predictions, images[idx], class_colors)
        batch_predictions.append(predictions)
    darknet.free_batch_detections(batch_detections, batch_size)
    return images, batch_predictions


def image_classification(image, network, class_names):
    width = darknet.network_width(network)
    height = darknet.network_height(network)
    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image_resized = cv2.resize(image_rgb, (width, height),
                                interpolation=cv2.INTER_LINEAR)
    darknet_image = darknet.make_image(width, height, 3)
    darknet.copy_image_from_bytes(darknet_image, image_resized.tobytes())
    detections = darknet.predict_image(network, darknet_image)
    predictions = [(name, detections[idx]) for idx, name in enumerate(class_names)]
    darknet.free_image(darknet_image)
    return sorted(predictions, key=lambda x: -x[1])


def convert2relative(image, bbox):
    """
    YOLO format use relative coordinates for annotation
    """
    x, y, w, h = bbox
    height, width, _ = image.shape
    return x/width, y/height, w/width, h/height


def save_annotations(name, image, detections, class_names):
    """
    Files saved with image_name.txt and relative coordinates
    """
    file_name = name.split(".")[:-1][0] + ".txt"
    with open(file_name, "w") as f:
        for label, confidence, bbox in detections:
            x, y, w, h = convert2relative(image, bbox)
            label = class_names.index(label)
            f.write("{} {:.4f} {:.4f} {:.4f} {:.4f}\n".format(label, x, y, w, h))


def batch_detection_example():
    args = parser()
    check_arguments_errors(args)
    batch_size = 3
    random.seed(3)  # deterministic bbox colors
    network, class_names, class_colors = darknet.load_network(
        args.config_file,
        args.data_file,
        args.weights,
        batch_size=batch_size
    )
    image_names = ['data/horses.jpg', 'data/horses.jpg', 'data/eagle.jpg']
    images = [cv2.imread(image) for image in image_names]
    images, detections,  = batch_detection(network, images, class_names,
                                           class_colors, batch_size=batch_size)
    for name, image in zip(image_names, images):
        cv2.imwrite(name.replace("data/", ""), image)
    print(detections)

def get_files(dir, suffix): 

    res = []

    for root, directory, files in os.walk(dir): 

        for filename in files:

            name, suf = os.path.splitext(filename) 

            if suf == suffix:

                #res.append(filename)

                res.append(os.path.join(root, filename))
    return res
def bbox2points_zs(bbox):
    """
    From bounding box yolo format
    to corner points cv2 rectangle
    """
    x, y, w, h = bbox
    xmin = int(round(x - (w / 2)))
    xmax = int(round(x + (w / 2)))
    ymin = int(round(y - (h / 2)))
    ymax = int(round(y + (h / 2)))
    return xmin, ymin, xmax, ymax

def main():
    args = parser()
    check_arguments_errors(args)
    input_dir = '/home/your/raid/darknet'
    config_file = '/home/your/raid/darknet/cfg/my_yolov3.cfg'
    data_file = '/home/your/darknet/cfg/my_data.data'
    weights = '/home/your/darknet/myData/backup/my_yolov3_last.weights'#修改为自己的路径
    random.seed(3)  # deterministic bbox colors
    network, class_names, class_colors = darknet.load_network(
        config_file,
        data_file,
        weights,
        batch_size=args.batch_size
    )
    src_width = darknet.network_width(network)
    src_height = darknet.network_height(network)

    #生成保存图片路径文件夹
    save_dir = os.path.join(input_dir, 'object_result')
    # 去除首位空格
    save_dir=save_dir.strip()
    # 去除尾部 \ 符号
    save_dir=save_dir.rstrip("\\")
    # 判断路径是否存在 # 存在     True # 不存在   False
    isExists=os.path.exists(save_dir)
    # 判断结果
    if not isExists:
        # 如果不存在则创建目录 # 创建目录操作函数
        os.makedirs(save_dir) 

        print(save_dir+' 创建成功')
    else:
        # 如果目录存在 则不创建，并提示目录已存在
        print(save_dir + ' 目录已存在')

    image_list = get_files(input_dir, '.jpg')
    total_len = len(image_list)
    index = 0
    #while True:
    for i in range(0, total_len):
        image_name = image_list[i]
        src_image = cv2.imread(image_name)
        cv2.imshow('src_image', src_image)
        cv2.waitKey(1)

        prev_time = time.time()
        image, detections = image_detection(
            image_name, network, class_names, class_colors, args.thresh)
        #'''
        file_name, type_name = os.path.splitext(image_name)
        #print(file_name)
        #print(file_name.split(r'/'))
        print(''.join(file_name.split(r'/')[-1]) + 'bbbbbbbbb')
        cut_image_name_list = file_name.split(r'/')[-1:] #cut_image_name_list is list
        save_dir_image = os.path.join(save_dir ,cut_image_name_list[0])
        if not os.path.exists(save_dir_image):
            os.makedirs(save_dir_image)
        cut_image_name = ''.join(cut_image_name_list) #list to str
        object_count = 0
        
        
        for label, confidence, bbox in detections:
            cut_image_name_temp = cut_image_name + "_{}.jpg".format(object_count)
            object_count += 1
            xmin, ymin, xmax, ymax = bbox2points_zs(bbox)
            print("aaaaaaaaa x,{} y,{} w,{} h{}".format(xmin, ymin, xmax, ymax))
            xmin_coordinary = (int)(xmin * src_image.shape[1] / src_width-0.5)
            ymin_coordinary = (int)(ymin * src_image.shape[0] / src_height-0.5)
            xmax_coordinary = (int)(xmax * src_image.shape[1] / src_width+0.5)
            ymax_coordinary = (int)(ymax * src_image.shape[0] / src_height+0.5)
            if xmin_coordinary>src_image.shape[1]:
                xmin_coordinary = src_image.shape[1]
            if ymin_coordinary>src_image.shape[0]:
                ymin_coordinary = src_image.shape[0]
            if xmax_coordinary>src_image.shape[1]:
                xmax_coordinary = src_image.shape[1]
            if ymax_coordinary>src_image.shape[0]:
                ymax_coordinary = src_image.shape[0]

            if xmin_coordinary < 0:
                xmin_coordinary = 0
            if ymin_coordinary < 0:
                ymin_coordinary = 0
            if xmax_coordinary < 0:
                xmax_coordinary = 0
            if ymax_coordinary < 0:
                ymax_coordinary = 0 

            print("qqqqqqqq   x,{} y,{} w,{} h{}".format(xmin_coordinary, ymin_coordinary, xmax_coordinary, ymax_coordinary))
            out_iou_img = np.full((ymax_coordinary - ymin_coordinary, xmax_coordinary - xmin_coordinary, src_image.shape[2]), 114, dtype=np.uint8)
            out_iou_img[:,:] = src_image[ymin_coordinary:ymax_coordinary,xmin_coordinary:xmax_coordinary]
            cv2.imwrite(os.path.join(save_dir_image,cut_image_name_temp),out_iou_img)
        #'''
        #if args.save_labels:
        #if True:
            #save_annotations(image_name, image, detections, class_names)
        darknet.print_detections(detections, args.ext_output)
        fps = int(1/(time.time() - prev_time))
        print("FPS: {}".format(fps))
        if not args.dont_show:
            #cv2.imshow('Inference', image)
            cv2.waitKey(1)
            #if cv2.waitKey() & 0xFF == ord('q'):
                #break
        index += 1

if __name__ == "__main__":
    # unconmment next line for an example of batch processing
    # batch_detection_example()
    main()

运行：

python3 detect.py

总结：以上就是我在这段时间的学习心得，主要目的是加深理解，也希望能帮到大家，如果有什么错误，也欢迎各位批评指正，共同进步！