darknet的实现

一、下载与安装

下载网络与权重文件&＃xff1a;点击打开链接

darkflow实现了将darknet翻译成tensorflow,可以用tensorflow加载darknet训练好的模型,并使用tensorflow重新训练,输出tensorflow graph模型,用于移动设备。dark flow需要安装的依赖库有numpy,tensorflow,opencv,python等。

打开下载好的文件夹&＃xff0c;里面有如下文件&＃xff1a;

[{&＃39;label&＃39;: &＃39;bear&＃39;, &＃39;confidence&＃39;: 0.33084026, &＃39;topleft&＃39;: {&＃39;x&＃39;: 186, &＃39;y&＃39;: 29}, &＃39;bottomright&＃39;: {&＃39;x&＃39;: 426, &＃39;y&＃39;: 380}}, {&＃39;label&＃39;: &＃39;dog&＃39;, &＃39;confidence&＃39;: 0.56702411, &＃39;topleft&＃39;: {&＃39;x&＃39;: 192, &＃39;y&＃39;: 58}, &＃39;bottomright&＃39;: {&＃39;x&＃39;: 418, &＃39;y&＃39;: 374}}]

可以看出&＃xff0c;结果还是比较准确的。

2.2 视频的测试

视频也是同样的&＃xff0c;放在同一文件夹下&＃xff0c;将图片换成视频&＃xff0c;一帧一帧的处理。

import cv2
import numpy as np
from darkflow.net.build import TFNet
import matplotlib.pyplot as pltoption &＃61; {&＃39;model&＃39;: &＃39;D:\Python\darkflow-master\cfg\yolo.cfg&＃39;, &＃39;load&＃39;: &＃39;D:/Python/darkflow-master/bin/yolov2.weights&＃39;,&＃39;thereshold&＃39;: 0.15}
tfnet &＃61; TFNet(option)
cap &＃61; cv2.VideoCapture("D:/Python/darkflow-master/test.mp4")
fps &＃61; 30
size &＃61; (1920, 1080)
writer &＃61; cv2.VideoWriter(&＃39;dark.avi&＃39;, -1, fps, size)
colors &＃61; [tuple(255 * np.random.rand(3)) for i in range(5)]
while (cap.isOpened()):ret, frame &＃61; cap.read()results &＃61; tfnet.return_predict(frame)if ret:for color, result in zip(colors, results):tl &＃61; (result[&＃39;topleft&＃39;][&＃39;x&＃39;], result[&＃39;topleft&＃39;][&＃39;y&＃39;])br &＃61; (result[&＃39;bottomright&＃39;][&＃39;x&＃39;], result[&＃39;bottomright&＃39;][&＃39;y&＃39;])label &＃61; result[&＃39;label&＃39;]cv2.rectangle(frame, tl, br, color, 7, )cv2.putText(frame, label, tl, cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 0), 2)writer.write(frame)cv2.imshow(&＃39;frame&＃39;,frame)if cv2.waitKey(1) & 0xFF &＃61;&＃61; ord(&＃39;q&＃39;):breakelse:cap.release()cv2.destroyAllWindows()break

由于不能上传视频&＃xff0c;只能截取其中一部分图片了

处理的结果如图&＃xff1a;

2.3  darkflow的配置文件

打开cfg文件下的yolo.cfg文件&＃xff0c;可以看到网络的构架如下&＃xff1a;

[net]
# Testing
batch&＃61;1
subdivisions&＃61;1
# Training
# batch&＃61;64
# subdivisions&＃61;8
width&＃61;608
height&＃61;608
channels&＃61;3
momentum&＃61;0.9
decay&＃61;0.0005
angle&＃61;0
saturation &＃61; 1.5
exposure &＃61; 1.5
hue&＃61;.1
learning_rate&＃61;0.001
burn_in&＃61;1000
max_batches &＃61; 500200
policy&＃61;steps
steps&＃61;400000,450000
scales&＃61;.1,.1
[convolutional]
batch_normalize&＃61;1
filters&＃61;32
size&＃61;3
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[maxpool]
size&＃61;2
stride&＃61;2
[convolutional]
batch_normalize&＃61;1
filters&＃61;64
size&＃61;3
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[maxpool]
size&＃61;2
stride&＃61;2
[convolutional]
batch_normalize&＃61;1
filters&＃61;128
size&＃61;3
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[convolutional]
batch_normalize&＃61;1
filters&＃61;64
size&＃61;1
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[convolutional]
batch_normalize&＃61;1
filters&＃61;128
size&＃61;3
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[maxpool]
size&＃61;2
stride&＃61;2
[convolutional]
batch_normalize&＃61;1
filters&＃61;256
size&＃61;3
stride&＃61;1
pad&＃61;1
activation&＃61;leaky
[convolutional]
batch_normalize&＃61;1
filters&＃61;128
size&＃61;1
stride&＃61;1
pad&＃61;1
activation&＃61;leaky

[convolutional]

batch_normalize&＃61;1

[convolutional]

batch_normalize&＃61;1

random&＃61;1

以上就是关于dark flow的全部内容&＃xff0c;更多详细内容见官网&＃xff1a;点击打开链接