Deep Residual Learning-based Reconstruction of Stacked Autoencoder Representation
Honggui Li, Maria Trocan · 2018
Stacked autoencoder (SAE) can efficiently represent high dimensional data with low dimensional features via minimizing a reconstruction error. However, the decoder of SAE cannot achieve lossless recovery of original data. If the reestablishment performance of SAE decoder is improved, it can be employed for low-bitrate and high-quality data compression. This paper adopts deep residual learning of convolutional neural network (CNN) to promote the reconstruction capability of SAE decoder. It is demonstrated by experimental results that deep CNN is superior to SAE decoder in rebuilding the performance of image data.