DnCNN

“Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, Lei Zhang, Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising, IEEE Trans. on Image Processing, 2017”

Network Architecture

 DnCNNs: feed-forward denoising convolutional neural networks

Code

def dncnn(input, is_training=True, output_channels=1):
    
    with tf.variable_scope('block1'):
        output = tf.layers.conv2d(input, 64, 3, padding='same', activation=tf.nn.relu)
         ## (i) Conv+ReLU: for the first layer
            # 64 filters of size 3*3*c are used to generate 64 feature maps
            # zero padding to avoid boundary artifacts
            # utilize rectified linearunits (ReLU, max(0; ·)) to introduce nonlinearity
            
    for layers in xrange(2, 16 + 1):
        with tf.variable_scope('block%d' % layers):
            output = tf.layers.conv2d(output, 64, 3, padding='same', name='conv%d' % layers, use_bias=False)
            output = tf.nn.relu(tf.layers.batch_normalization(output, training=is_training))
        ## (ii) Conv+BN+ReLU: for layers 2 ~ (D - 1)
            # 64 filters of size 3*3*64 
            # batch normalization is added between convolution and ReLU  
            
    with tf.variable_scope('block17'):
        output = tf.layers.conv2d(output, output_channels, 3, padding='same')
        ## (iii) Conv: for the last layer
            # c filters of size 3*3*64 are used to reconstruct the output
            
     ## output is residual image; return is (input-output) = predicted clean image       
    return input - output

网络结构说明：

1. First layer：Conv（3 * 3 * c * 64）+ReLu （c代表图片通道数）
2. Layers 2~(D-1)：Conv（3 * 3 * 64 * 64）+BN(batch normalization)+ReLu
3. Last layer：Conv（3 * 3 * 64）
4. 每一层都zero padding，使得每一层的输入、输出尺寸保持一致，防止产生 boundary artifact。

Highlights

1. Residual learning：解决网络深度加深带来的梯度消失或梯度爆炸问题☞ResNet论文详解
2. Batch normalization：解决内部协变量移位问题☞ BN生动形象的说明

Experimental setting

1. 400 images of size 180×180 for training
2. three noise levels, i.e., σ = 15, 25 and 50
3. patch size = 40 × 40, crop 128 × 1600 patches to train the model
4. network depth = 17
5. mini-batch size = 128
6. training epochs = 50
7. learning rate was decayed exponentially from 1e-1 to 1e-4 for the 50 epochs.