"Paper notes" Object contour detection with a fully convolutional encoder-decoder network

Object Contour Detection with a fully convolutional encoder-decoder network

Using convolutional encoding and decoding network to detect the edges of primary targets

The network structure is:
Code: VGG-16
Decoding: Reverse pooling-convolution-activation-dropout

Convolution cores:

The number of channels of every decoder layer is properly
Designed to allow unpooling the maxpooling layer from its corresponding.

Dropout:

We also add a dropout layer after each relu layer in the decoder. A Complete Decoder Network setup is listed in Table 1

Loss function:

The loss function is simply the Pixel-wise logistic loss.

Anti-pool:

We Initialize our encoder with VGG-16 net [+] (up to the "fc6" layer) and to achieve dense prediction of image size our D Ecoder is constructed by alternating unpooling and convolution layers where unpooling layers re-use the switches from Max-pooling layers of encoder to upscale the feature maps.

Counter-pooling Details:
http://blog.csdn.net/xiaxzhou/article/details/74012137

Training:
During training, we fix the encoder parameters (VGG-16) and only optimize decoder.