Deconvolution neural network

lastmod 2026-11-10created 2017-06-273 edits

Deconvolution network (deconvolution neural network) is a neural network architecture that expands feature maps shrunk by convolution back to the original resolution. It stacks unpooling and transposed convolution to form an encoder-decoder structure, and was proposed to improve the coarse output of FCN in semantic segmentation, which requires pixel-level prediction. Despite its name, it is not a mathematical deconvolution but a learnable upsampling layer.

Documentation

Learning Deconvolution Network for Semantic Segmentation

http://cvlab.postech.ac.kr/research/deconvnet/

https://arxiv.org/abs/1505.04366

See also

• Machine learning
• CNN