1. Anti-pooling and deconvolution
The following is the overall architecture of DeconvNet:
unpooling：
(The Unpooling operation needs to restore each pooling area based on the position recorded by max pooling. When performing max pooling, we record the coordinates of the position where the maximum activation value of each pooling area is located. During unpooling, these coordinates can be used. Put the activation value back to the corresponding position to achieve the purpose of unpooling. The purpose of Unpooling is to perform upsampling and restore a higher resolution feature map.)
Deconvolution
Deconvolution simply transforms the input back to a larger size. In transposed convolution, you try to upscale some features to approximate the image. Transposed convolutions have learnable properties and can be learned using backpropagation.
A deconvolutional layer is a very unfortunate name, it should be called a transposed convolutional layer.
The process of deconvolving an input image with a filter. Deconvolution is an operation used for upsampling, that is, converting a low-resolution input to a high-resolution output. Deconvolution is not the inverse process of convolution, but a special convolution whose convolution kernel is transposed and inserts some zeros into the input while changing the step size to 1, 1/2, 1 /3 or 1/4 to increase the area of the feature map.
Visually, for a transposed convolution with stride one and no padding, we just pad the original input (blue entries) with zeroes (white entries) (Figure 1).
In case of stride two and padding, the transposed convolution would look like this (Figure 2):
The picture above is an example. (b) is the output of the 14×14 deconvolution layer. (c) is the output after unpooling, and so on. We can see in (j) that bicycles can be reconstructed at the last 224×224 deconvolutional layer, suggesting that the learned filters can capture class-specific shape information.