图像语义分割——抠图勾画

1. 目前所在公司的核心产品，是放疗图像的靶区自动勾画。

2. 使用深度学习技术，学习放疗样本，能够针对不同的器官，进行放疗靶区的勾画。

3. 使用CNN搭建FCN/U-Net网络结构，训练模型，使模型获得图像语义分隔的能力。（自动驾驶，无人机落点判定都是属于语义分割范畴）。

FCN模型结构

模型结构如图所示

模型流程

1. 对输入图片image，重复进行CNN卷积、激活函数和池化等处理，得到pool1,pool2,pool3,pool,poo5，5个池化结果
2. 将pool5池化结果进行反卷积放大2倍，得到结果A
3. 将A与pool4相加得到结果B
4. 将结果B进行反卷积放大2倍，得到结果C
5. 将结果C与pool3相加得到结果D，
6. 将结果D经过激活函数处理，得到最终的图像result

模型代码

weight = {
    'w1sub': tf.Variable(tf.random_normal([8, 8, 1, 8], stddev=0.1)),
    'w2sub': tf.Variable(tf.random_normal([4, 4, 8, 16], stddev=0.1)),
    'w3sub': tf.Variable(tf.random_normal([2, 2, 16, 32], stddev=0.1)),
    'w4sub': tf.Variable(tf.random_normal([2, 2, 32, 64], stddev=0.1)),

    'w1up': tf.Variable(tf.random_normal([2, 2, 32, 64], stddev=0.1)),
    'w2up': tf.Variable(tf.random_normal([2, 2, 16, 32], stddev=0.1)),
    'w3up': tf.Variable(tf.random_normal([2, 2, 8, 16], stddev=0.1)),
    'w4up': tf.Variable(tf.random_normal([2, 2, 1, 8], stddev=0.1)),
}

biases = {
    'b1sub': tf.Variable(tf.random_normal([8], stddev=0.1)),
    'b2sub': tf.Variable(tf.random_normal([16], stddev=0.1)),
    'b3sub': tf.Variable(tf.random_normal([32], stddev=0.1)),
    'b4sub': tf.Variable(tf.random_normal([64], stddev=0.1)),

    'b1up': tf.Variable(tf.random_normal([32], stddev=0.1)),
    'b2up': tf.Variable(tf.random_normal([16], stddev=0.1)),
    'b3up': tf.Variable(tf.random_normal([8], stddev=0.1)),
    'b4up': tf.Variable(tf.random_normal([1], stddev=0.1)),
}


def ForwardProcess(inputBatch, w, b, num_size, istrain=False):
    inputBatch_r = tf.reshape(inputBatch, shape=[-1, 400, 400, 1])
    if istrain:
    #dropout处理
        inputBatch_r = tf.nn.dropout(inputBatch_r, keep_prob=0.9)

    conv1 = tf.nn.conv2d(inputBatch_r, w['w1sub'], strides=[1, 1, 1, 1], padding='SAME')  # 8
    conv1 = tf.layers.batch_normalization(conv1, training=True)
    conv1 = tf.nn.relu(tf.nn.bias_add(conv1, b['b1sub']))
    pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')  # 8
    print('pool1的形状')
    print(pool1.get_shape())

    conv2 = tf.nn.conv2d(pool1, w['w2sub'], strides=[1, 1, 1, 1], padding='SAME')  # 16
    conv2 = tf.layers.batch_normalization(conv2, training=True)
    conv2 = tf.nn.relu(tf.nn.bias_add(conv2, b['b2sub']))
    pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    print('pool2的形状')
    print(pool2.get_shape())

    conv3 = tf.nn.conv2d(pool2, w['w3sub'], strides=[1, 1, 1, 1], padding='SAME')  # 32
    conv3 = tf.layers.batch_normalization(conv3, training=True)
    conv3 = tf.nn.relu(tf.nn.bias_add(conv3, b['b3sub']))
    pool3 = tf.nn.max_pool(conv3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    print('pool3的形状')
    print(pool3.get_shape())

    conv4 = tf.nn.conv2d(pool3, w['w4sub'], strides=[1, 1, 1, 1], padding='SAME')  # 64
    conv4 = tf.layers.batch_normalization(conv4, training=True)
    conv4 = tf.nn.relu(tf.nn.bias_add(conv4, b['b4sub']))
    pool4 = tf.nn.max_pool(conv4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
    print('pool4的形状')
    print(pool4.get_shape())

    # 反卷积

    print('第一次反卷积，输入的形状')
    print(pool4.get_shape())
    d_conv1 = tf.nn.conv2d_transpose(value=pool4, filter=w['w1up'], output_shape=[num_size, 50, 50, 32],
                                     strides=[1, 2, 2, 1], padding='SAME')
    d_conv1 = tf.add(d_conv1, pool3)
    d_conv1 = tf.nn.bias_add(d_conv1, b['b1up'])
    d_conv1 = tf.layers.batch_normalization(d_conv1, training=True)
    d_conv1 = tf.nn.relu(d_conv1)

    print('第二次反卷积，输入的形状')
    print(d_conv1.get_shape())
    d_conv2 = tf.nn.conv2d_transpose(value=d_conv1, filter=w['w2up'], output_shape=[num_size, 100, 100, 16],
                                     strides=[1, 2, 2, 1], padding='SAME')
    d_conv2 = tf.add(d_conv2, pool2)
    d_conv2 = tf.nn.bias_add(d_conv2, b['b2up'])
    d_conv2 = tf.layers.batch_normalization(d_conv2, training=True)
    d_conv2 = tf.nn.relu(d_conv2)

    print('第三次反卷积，输入的形状')
    print(d_conv2.get_shape())
    d_conv3 = tf.nn.conv2d_transpose(value=d_conv2, filter=w['w3up'], output_shape=[num_size, 200, 200, 8],
                                     strides=[1, 2, 2, 1], padding='SAME')
    d_conv3 = tf.add(d_conv3, pool1)
    d_conv3 = tf.nn.bias_add(d_conv3, b['b3up'])
    d_conv3 = tf.layers.batch_normalization(d_conv3, training=True)
    d_conv3 = tf.nn.relu(d_conv3)

    print('第四次反卷积，输入的形状')
    print(d_conv3.get_shape())
    d_conv4 = tf.nn.conv2d_transpose(value=d_conv3, filter=w['w4up'], output_shape=[num_size, 400, 400, 1],
                                     strides=[1, 2, 2, 1], padding='SAME')
    # d_conv4 = tf.add(d_conv4, inputBatch_r)
    d_conv4 = tf.nn.bias_add(d_conv4, b['b4up'])
    d_conv4 = tf.layers.batch_normalization(d_conv4, training=True)
    d_conv4 = tf.nn.relu(d_conv4)

    return d_conv4

未完待续