Parameterization of a Convolutional Autoencoder for Reconstruction of Small Images

Tang, L. M., Lim, L. H. I. and Siebert, P. (2018) Parameterization of a Convolutional Autoencoder for Reconstruction of Small Images. In: 15th International Conference on Control, Automation, Robotics and Vision (ICARCV 2018), Singapore, 18-21 Nov 2018, (Accepted for Publication)

[img] Text
170780.pdf - Accepted Version
Restricted to Repository staff only



A convolutional autoencoder (CAE) is formed by combining a convolutional neural network and an autoencoder, to take both their advantages in reconstructing the output from a compact, latent representation of the input. However, to our best knowledge, there is no exact recommendation for parameterizing a CAE, such as deciding the number of neurons in the hidden bottleneck layer of a CAE to avoid ”underfitting” and ”overfitting” of the network. Hence, a framework for deriving an optimum set of CAE parameters for the reconstruction of input images based on the standard MNIST data set is presented in this paper. The robustness of the parameters on a different image size’s data set, like the SVHN, is then verified. Our results show that for small (28 x 28) and (32 x 32) pixels’ input images, having 2560 neurons at the hidden bottleneck layer and 32 convolutional feature maps can result in optimum reconstruction performance for the CAEs. In addition, the quantitative Mean-Square-Error and the qualitative (2D visualization of the neurons’ activation, the histogram statistics and estimated source entropy at the hidden layers) analysis methodology provided by this work can provide a good framework for deciding the parameter values of the CAEs to provide good representations of the input image.

Item Type:Conference Proceedings
Status:Accepted for Publication
Glasgow Author(s) Enlighten ID:Siebert, Dr Jan and Tang, Lai Meng and Lim, Dr Li Hong Idris
Authors: Tang, L. M., Lim, L. H. I., and Siebert, P.
Subjects:Q Science > Q Science (General)
College/School:College of Science and Engineering > School of Computing Science
College of Science and Engineering > School of Engineering > Electronics and Nanoscale Engineering
Related URLs:

University Staff: Request a correction | Enlighten Editors: Update this record