您的位置: 首页  >  IT文章  >  目标检测算法-YOLO-V1训练代码详解

目标检测算法-YOLO-V1训练代码详解

分类: IT文章 • 2024-09-30 08:22:31

YOLO-V1网络结构由24个卷积层与2个全连接层构成，网络入口为448×448×3，输出维度：S×S×(B×5+C)，S为划分网格数，B为每个网格负责目标个数，C为类别个数。

目标检测算法-YOLO-V1训练代码详解

YOLO-V1是将一副图像分成S×S个网格，如果某个object的中心落在这个网格中，则这个网格就负责预测这个object，每个网格要预测B个bounding box，每个bounding box要预测一个confidence值，这个confidence代表了所预测的bounding box中含有object的置信度和这个bounding box预测的有多准这两个重要信息。

Pr(Object)∗IoUpredtruth​

如果有object落在一个网格中，公式第一项取1，否则取0，第二项是bounding box和真实框的IOU的值(confidence针对每个bounding box，框中有没有网格包含object中心点。YOLO-V1中每个网格有两个bounding box，对于每个bounding box有5个预测值，x,y,w,h,confidence,每一个网格还要预测C条件类别的概率，即在一个网格包含一个object的前提下，它属于某个类别的概率。(x,y)表示bounding box相对于网格单元的边界的offset，归一化到(0,1)范围之内，而w,h表示相对于整个图片的预测宽和高，也被归一化到(0,1)范围内。c代表的是object在某个bounding box的confidence。confidence计算如下：

 Pr(Classi​∣Object)∗Pr(Object)∗IoUpredtruth​=Pr(Classi​)∗IoUpredtruth​

下面说明如何将预测坐标的x,y用相对于对应网格的offset归一化到0-1和w,h是如何利用图像的宽高归一化到0-1之间。每个单元格预测的B个(x,y,w,h,confidence)向量，假设图片为S×S个网格，S=7，图片宽为w​i高为hi 。

下面引用一张我看过的感觉讲解很详细的一张图片：

目标检测算法-YOLO-V1训练代码详解

​

 ​

YOLO-V1中需要的参数

 1 def __init__(self):
 2     self.classes = ["aeroplane", "bicycle", "bird", "boat", "bottle",
 3                     "bus", "car", "cat", "chair", "cow", "diningtable",
 4                     "dog", "horse", "motorbike", "person", "pottedplant",
 5                     "sheep", "sofa", "train", "tvmonitor"]
 6     #计算坐标用的
 7     self.x_offset = np.transpose(np.reshape(np.array([np.arange(7)] * 7 * 2, dtype=np.float32), [2, 7, 7]), [1, 2, 0])
 8     self.y_offset = np.transpose(self.x_offset, [1, 0, 2])
 9     #输入图片大小
10     self.img_size = (448, 448)
11     #阈值
12     self.iou_threshold = 0.5
13     self.batch_size = 45
14     #计算loss需要的参数
15     self.class_scale = 2.0
16     self.object_scale = 1.0
17     self.noobject_scale = 1.0
18     self.coord_scale = 5.0

网络部分开始

 1 def _build_net(self):
 2     x = tf.placeholder(tf.float32, [None, 448, 448, 3])
 3     with tf.variable_scope('yolo'):
 4         net = self.conv_layer(x, 64, 7, 2, 'conv_2')
 5         net = self.max_pool_layer(net, 2, 2)
 6         net = self.conv_layer(net, 192, 3, 1, 'conv_4')
 7         net = self.max_pool_layer(net, 2, 2)
 8         net = self.conv_layer(net, 128, 1, 1, 'conv_6')
 9         net = self.conv_layer(net, 256, 3, 1, 'conv_7')
10         net = self.conv_layer(net, 256, 1, 1, 'conv_8')
11         net = self.conv_layer(net, 512, 3, 1, 'conv_9')
12         net = self.max_pool_layer(net, 2, 2)
13         net = self.conv_layer(net, 256, 1, 1, 'conv_11')
14         net = self.conv_layer(net, 512, 3, 1, 'conv_12')
15         net = self.conv_layer(net, 256, 1, 1, 'conv_13')
16         net = self.conv_layer(net, 512, 3, 1, 'conv_14')
17         net = self.conv_layer(net, 256, 1, 1, 'conv_15')
18         net = self.conv_layer(net, 512, 3, 1, 'conv_16')
19         net = self.conv_layer(net, 256, 1, 1, 'conv_17')
20         net = self.conv_layer(net, 512, 3, 1, 'conv_18')
21         net = self.conv_layer(net, 512, 1, 1, 'conv_19')
22         net = self.conv_layer(net, 1024, 3, 1, 'conv_20')
23         net = self.max_pool_layer(net, 2, 2)
24         net = self.conv_layer(net, 512, 1, 1, 'conv_22')
25         net = self.conv_layer(net, 1024, 3, 1, 'conv_23')
26         net = self.conv_layer(net, 512, 1, 1, 'conv_24')
27         net = self.conv_layer(net, 1024, 3, 1, 'conv_25')
28         net = self.conv_layer(net, 1024, 3, 1, 'conv_26')
29         net = self.conv_layer(net, 1024, 3, 2, 'conv_28')
30         net = self.conv_layer(net, 1024, 3, 1, 'conv_29')
31         net = self.conv_layer(net, 1024, 3, 1, 'conv_30')
32         net = self.flatten_layer(net)
33         net = self.dense_layer(net, 512, activation=self.Leaky_Relu, scope='fc_33')
34         net = self.dense_layer(net, 4096, activation=self.Leaky_Relu, scope='fc_34')
35         net = self.dense_layer(net, 7 * 7 * 30, scope='fc_36')
36     return net

需要的一些层

 1 # 激活函数使用Leaky
 2 def Leaky_Relu(self, x):
 3     return tf.maximum(x * 0.1, x)
 4 # 卷积层
 5 def conv_layer(self, x, filter, kernel_size, stride, scope):
 6     channel = x.get_shape().as_list()[-1]
 7     weight = tf.Variable(tf.truncated_normal(shape=[kernel_size, kernel_size, channel, filter], stddev=0.1),
 8                          name="weights")
 9     bias = tf.Variable(tf.zeros([filter, ]), name="biases")
10     pad_size = kernel_size // 2
11     x = tf.pad(x, paddings=[[0, 0], [pad_size, pad_size], [pad_size, pad_size], [0, 0]])
12 
13     conv = tf.nn.conv2d(x, weight, strides=[1, stride, stride, 1], padding="VALID", name=scope)
14     output = self.Leaky_Relu(tf.nn.bias_add(conv, bias))
15     return output
16 # 最大池化层
17 def max_pool_layer(self, x, pool_size, stride):
18     return tf.nn.max_pool(x, [1, pool_size, pool_size, 1], strides=[1, stride, stride, 1], padding="SAME")
19 # 全连接层
20 def dense_layer(self, x, filter, activation=None, scope=None):
21     channel = x.get_shape().as_list()[-1]
22     weight = tf.Variable(tf.truncated_normal(shape=[channel, filter], stddev=0.1), name="weights")
23     bias = tf.Variable(tf.zeros([filter, ]), name="biases")
24     output = tf.nn.xw_plus_b(x, weight, bias, name=scope)
25     if activation:
26         output = activation(output)
27     return output
28 # flatten层
29 def flatten_layer(self, x):
30     x = tf.transpose(x, [0, 3, 1, 2])
31     shape = x.get_shape().as_list()[1:]
32     nums = np.product(shape)
33     return tf.reshape(x, [-1, nums])

网络部分结束

相关推荐