FCN

最新推荐文章于 2025-02-25 20:48:50 发布

comeontoto

最新推荐文章于 2025-02-25 20:48:50 发布

阅读量5.3k

点赞数 2

分类专栏：网络文章标签： FCN

本文链接：https://blog.csdn.net/comeontoto/article/details/79954337

版权

网络专栏收录该内容

1 篇文章

订阅专栏

来源：https://blog.csdn.net/zhyj3038/article/details/71195262

1 实例分割,

对图像中的每一个像素点进行分类,同种物体的不同实例也用不同的类标进行标注，中间是实例分割,右图是语义分割.

2FCN与CNN区别

3CNN中Alexnet的代码：https://blog.csdn.net/akadiao/article/details/78592290?locationNum=10&fps=1

def AlexNet(images, classNum=None, dropoutrate=None):
    parameters = []

    # 卷积层1
    conv1, parameters = convLayer(images, name='conv1', kh=11, kw=11, n_out=64, dh=4, dw=4, p=parameters)
    # 添加LRN层和最大池化层
    # 对conv1进行LRN处理
    lrn1 = tf.nn.lrn(conv1, 4, bias=1.0, alpha=0.001/9, beta=0.75, name='lrn1')
    # 对lrn1进行最大池化处理,池化尺寸3*3,步长2*2,padding模式选VALID即取样不能超过边框
    pool1 = tf.nn.max_pool(lrn1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool1')
    # 打印出结果pool1的结构
    print pool1.op.name, ' ', pool1.get_shape().as_list()

    # 卷积层2
    conv2, parameters = convLayer(pool1, name='conv2', kh=5, kw=5, n_out=192, dh=1, dw=1, p=parameters)
    # LRN处理
    lrn2 = tf.nn.lrn(conv2, 4, bias=1.0,alpha=0.001/9, beta=0.75, name='lrn2')
    # 最大池化处理
    pool2 = tf.nn.max_pool(lrn2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool2')
    # 打印出结果pool2的结构
    print pool2.op.name, ' ', pool2.get_shape().as_list()

    # 卷积层3
    conv3, parameters = convLayer(pool2, name='conv3', kh=3, kw=3, n_out=384, dh=1, dw=1, p=parameters)

    # 卷积层4
    conv4, parameters = convLayer(conv3, name='conv4', kh=3, kw=3, n_out=256, dh=1, dw=1, p=parameters)

    # 卷积层5
    conv5, parameters = convLayer(conv4, name='conv5', kh=3, kw=3, n_out=256, dh=1, dw=1, p=parameters)
    pool5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name='pool5')
    print pool5.op.name, ' ', pool5.get_shape().as_list()

    fc_in = tf.reshape(pool5, [-1, 256*6*6])
    fc6 = fcLayer(fc_in, 256*6*6, 4096, True, 'fc6')
    dropout6 = tf.nn.dropout(fc6, dropoutrate)

    fc7 = fcLayer(dropout6, 4096, 4096,True, 'fc7')
    dropout7 = tf.nn.dropout(fc7, dropoutrate)

    fc8 = fcLayer(dropout7, 4096, classNum, True, 'fc8')

    return pool5, parameters

FCN网络结构的部分代码，将三个全连接换成了卷积层：

with tf.variable_scope("inference"):
    image_net = vgg_net(weights, processed_image)
    conv_final_layer = image_net["conv5_3"]

    pool5 = utils.max_pool_2x2(conv_final_layer)

    W6 = utils.weight_variable([7, 7, 512, 4096], name="W6")
    b6 = utils.bias_variable([4096], name="b6")
    conv6 = utils.conv2d_basic(pool5, W6, b6)
    relu6 = tf.nn.relu(conv6, name="relu6")
    if FLAGS.debug:
        utils.add_activation_summary(relu6)
        #根据给出的keep_prob参数，将输入tensor x按比例输出。
    relu_dropout6 = tf.nn.dropout(relu6, keep_prob=keep_prob)

    W7 = utils.weight_variable([1, 1, 4096, 4096], name="W7")#图片的尺寸变成1*1，卷积核的数量4096，输出的尺寸就是1*1*4096，最后额
    b7 = utils.bias_variable([4096], name="b7")
    conv7 = utils.conv2d_basic(relu_dropout6, W7, b7)
    relu7 = tf.nn.relu(conv7, name="relu7")
    if FLAGS.debug:
        utils.add_activation_summary(relu7)
    relu_dropout7 = tf.nn.dropout(relu7, keep_prob=keep_prob)

    W8 = utils.weight_variable([1, 1, 4096, NUM_OF_CLASSESS], name="W8")
    b8 = utils.bias_variable([NUM_OF_CLASSESS], name="b8")
    conv8 = utils.conv2d_basic(relu_dropout7, W8, b8)
    # annotation_pred1 = tf.argmax(conv8, dimension=3, name="prediction1")

    # now to upscale to actual image size 现在升级为实际的图像大小
    deconv_shape1 = image_net["pool4"].get_shape()
    W_t1 = utils.weight_variable([4, 4, deconv_shape1[3].value, NUM_OF_CLASSESS], name="W_t1")
    b_t1 = utils.bias_variable([deconv_shape1[3].value], name="b_t1")
    conv_t1 = utils.conv2d_transpose_strided(conv8, W_t1, b_t1, output_shape=tf.shape(image_net["pool4"]))
    fuse_1 = tf.add(conv_t1, image_net["pool4"], name="fuse_1")

    deconv_shape2 = image_net["pool3"].get_shape()
    W_t2 = utils.weight_variable([4, 4, deconv_shape2[3].value, deconv_shape1[3].value], name="W_t2")
    b_t2 = utils.bias_variable([deconv_shape2[3].value], name="b_t2")
    conv_t2 = utils.conv2d_transpose_strided(fuse_1, W_t2, b_t2, output_shape=tf.shape(image_net["pool3"]))
    fuse_2 = tf.add(conv_t2, image_net["pool3"], name="fuse_2")

    shape = tf.shape(image)
    deconv_shape3 = tf.stack([shape[0], shape[1], shape[2], NUM_OF_CLASSESS])
    W_t3 = utils.weight_variable([16, 16, NUM_OF_CLASSESS, deconv_shape2[3].value], name="W_t3")
    b_t3 = utils.bias_variable([NUM_OF_CLASSESS], name="b_t3")
    conv_t3 = utils.conv2d_transpose_strided(fuse_2, W_t3, b_t3, output_shape=deconv_shape3, stride=8)

    annotation_pred = tf.argmax(conv_t3, dimension=3, name="prediction")

return tf.expand_dims(annotation_pred, dim=3), conv_t3
#对conv8此时进行32倍的上采样可以得到原图大小,这个时候得到的结果就是叫做FCN-32s. 
#在FCN-32s的基础上进行fine tuning,把pool4层和conv8的2倍上采样结果相加之后进行一个16倍的上采样,得到的结果是FCN-16s.如conv_t1
#FCN-16s的基础上进行fine tuning,把pool3层和2倍上采样的pool4层和4倍上采样的conv8层加起来,进行一个8倍的上采样,得到的结果就是FCN-8s