Abstract:

The paper describes the main problem associated with the use of convolutional networks in digital image processing - the loss of some information during subsampling operations performed by convolutional and pooling layers. The reverse operation to convolution is presented - transposed convolution, which is one of the methods that allow reconstruction of the original dimensions of the input image processed by convolutional layers. The autoencoder architecture is also described, and the main areas of computer vision using the transformations described in the paper are defined.