Deep Learning: deconvolution (transposed convolution) introduction

inverse convolution and convolution

Deconvolution, as the name suggests, is the reverse operation of convolution operations.

In order to facilitate understanding, suppose convolution is a picture before convolution is the characteristic of the picture.

Convolution, input picture, output picture characteristics, theoretical basis is the statistical invariance of Translational invariance (translation invariance), play a role in dimensionality reduction. Move diagram as follows:

Deconvolution, input the characteristics of the picture, output pictures, play a role in reducing. Move diagram as follows:

We know that convolution combined with the reduced-dimensional role of the pool, the neural network to the depth of learning, open the new World of neural networks, then the role of deconvolution. the effect of deconvolution

The deconvolution emerges in "unsupervised representations Learning with Deep convolutionalgenerative adversarial Networks", The main work of this paper is to use Gans to generate pictures, of which generator and discriminator are used in depth learning, Generator to generate pictures in the process of using deconvolution operations (of course, discriminator to use the convolution of the generator generated by the image to distinguish the authenticity). Generator generate pictures to indicate:
types of Reverse convolution

General type:

Input: 2*2
Calculation process: 2+2*2-3+1
Output: 4

Fractionally-strided convolution:

Input: 3*3
Calculation process: 3+4-3+1
Output: 5

Fractionally-strided, the number of decimal steps, reflected in the gap padding, nominally stride=1, but the actual effect is less than 1. the realization of deconvolution in TensorFlow

Tf.nn.conv2d_transpose (value, filter, Output_shape, strides, padding= ' SAME ', data_format= ' NHWC ', Name=none)
Parameter explanation:

Value:[batch,in_height,in_width,in_channels]
Filter:[batch,k_width,out_channels,in_channels]
Output_shape:[batch,output_height,output_width,output_channels]
Strides: The number of moves that correspond to each dimension