Densely Connected convolutional Networks paper Reading

Bishi finally came to an end, the traditional method of vision to do my whole people are very run, finally ended, you can see some of the papers on hold for a long time, boingonium hindering

Densely Connected convolutional Networks actually came out early, CVPR best paper

Think before reading the paper, or the entire network structure of dense net to Http://ethereon.github.io/netscope/#/editor above the visual look, it will be easier to understand, the overall paper is very good understanding

is a 5-storey dense block, each dense block of growth rate k=4

Several advantages of densnet are given at the beginning of the thesis:

1, our proposed densenet architecture explicitly differentiates between information, that's added to the network and inform ation is preserved. Densenet layers is very narrow (e.g., filters per layer), adding only a small set of feature-maps to the "collective kn Owledge "of the network and keep the remaining featuremaps
Unchanged-and The final classifier makes a decision based on all feature-maps in the network

DENSNET network structure parameters are few, each block within the filter is also relatively small, and we are using alexnet, usually filter is hundreds of, and here the filter 12, 24, 16, etc., so very narrow

2. One big advantage of densenets is their improved flow of information and gradients throughout the network, which makes T Hem easy to train.

Densenet network from which can be seen, each layer is connected with the back layer, (the first picture does not draw the connection between the layers in each block, it should be combined with the first diagram and the second figure, only to calculate a complete, Because the second figure, the input behind each block is the result of all the previous layers concat together, as shown in Figure one .... In the visual tools, the most obvious, but also can see the actual size of each layer is conducive to information and gradients throughout the network transmission.

3, we also observe that dense connections has a regularizing effect, which reduces overfitting on tasks with smaller train ing set sizes.

At the same time, Densenet network also has the function of regularization, training on small data sets can reduce the risk of overfitting.

Densenet is the concat of the previous layers, and ResNet is the summation, which is mentioned in the paper, which affects the transmission of information in the network.

Transition layers:

This layer is the layer connected between the two blocks, consisting of the BN layer, the 1x1 convolution layer, and the 2x2 AVG pooling layer, as shown in

Growth rate:

That is, the number of layers inside each block, one in which there are 4 layers in each block, so growth rate=4

Botttleneck layers:

It has been noted in [$, one] that a 1x1 convolution can is introduced as bottleneck layer before each 3x3 convolution to Reduce the number of input feature-maps, and thus to improve computational efficiency.

Is that there is a 1x1 convolution core in front of the 3x3 convolution core in each layer, which reduces the number of input feature-map, as shown, 512 of the number becomes 128

Compression:

If a dense block contains m feature-maps, we let the following transition layer generate [θm] output Featuremaps,where 0& Lt;θ<=1referred to as the compression factor.

Let the transition layer compress the number of feature maps for the block output.

Densely Connected convolutional Networks paper read