A mixed-scale dense convolutional neural network for image analysis
Published in PNAS on December 26, 2017
Available at PNAS Online: https://doi.org/10.1073/pnas.1715832114
Danie ̈l M. Pelt and James A. Sethian
Written before: The method in this paper cannot be implemented using existing frameworks such as TensorFlow or Caffe.
To briefly summarize:
contribute:
We propose a new neural network (based on atrous convolutions and dense connections) that can achieve better results on segmentation tasks with a network that is better trained with fewer parameters.
detail:
In essence, it is still pixel-pixel segmentation, but there is no upsampling process.
Different channels in each layer use different dilations.
From input to output, the size of each layer is the same, which is convenient for dense connection, which means that all channels of all previous layers can be used in the current operation. The authors argue that such processing maximizes the reuse of existing feature maps.
All layers use 3*3 atrous convolution, and the last layer uses 1*1 convolution (equivalent to the last layer being a linear combination of all channels of all previous layers).
The number of channels of each layer is the same as w, and the number of hidden layers is assumed to be d. The author exemplifies the network connection method, as shown in the following figure:
advantage:
Fast training, few parameters, and low risk of overfitting.
shortcoming:
There is no way to quickly build implementations using existing frameworks.