Presentation | 2018-03-18 Learning Convolutional Autoencoders Using a Loss Function Based on Spatial Frequencies and Colors Naoyuki Ichimura, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | This paper presents a learning method for convolutional autoencoders (CAEs) for extracting features from images. CAEs can be obtained by utilizing convolutional neural networks to learn an approximation to the identity function in an unsupervised manner. The loss function based on the pixel loss (PL) that is computed from the mean squared errors between the pixel values of original images and reconstructed images is the common choice for learning. However, using the loss function leads to blurred reconstructed images and requires a large number of epochs to reproduce colors. A method for learning CAEs using a loss function based on spatial frequencies and colors is proposed to mitigate the problems. The blurs in reconstructed images show lack of high spatial frequency components. In order to evaluate the lack of components, a convolutional layer with a Laplacian filter bank as weights is added to CAEs and the mean squared error of a subband (Spatial Frequency Loss:SFL) is obtained from the output of each filter. The mean squared error of a chromatic component (Chromatic Loss:CL) is introduced as well to evaluate color reproduction by addition a convolutional layer by which chromatic components are separated from a luminance component. The learning is performed using a loss function based on the SFL and CL. Empirical evaluation demonstrates that using the SFL reduces the blurs and using the CL facilitates the reproduction of colors. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Convolutional autoencoders / Unsupervised learning / Loss function / Spatial frequency / Color |
Paper # | BioX2017-36,PRMU2017-172 |
Date of Issue | 2018-03-11 (BioX, PRMU) |
Conference Information | |
Committee | PRMU / BioX |
---|---|
Conference Date | 2018/3/18(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | Shinichi Sato(NII) / Kazuhiko Sumi(AGU) |
Vice Chair | Hironobu Fujiyoshi(Chubu Univ.) / Yoshihisa Ijiri(Omron) / Hiroshi Takano(Toyama Pref. Univ.) / Hitoshi Imaoka(NEC) |
Secretary | Hironobu Fujiyoshi(AIST) / Yoshihisa Ijiri(NAIST) / Hiroshi Takano(Shizuoka Univ.) / Hitoshi Imaoka(Fujitsu Labs.) |
Assistant | Masato Ishii(NEC) / Yusuke Sugano(Osaka Univ.) / Masatsugu Ichino(Univ. of Electro-Comm.) / Naoyuki Takada(Secom) / Norihiro Okui(KDDI Research) |
Paper Information | |
Registration To | Technical Committee on Pattern Recognition and Media Understanding / Technical Committee on Biometrics |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Learning Convolutional Autoencoders Using a Loss Function Based on Spatial Frequencies and Colors |
Sub Title (in English) | |
Keyword(1) | Convolutional autoencoders |
Keyword(2) | Unsupervised learning |
Keyword(3) | Loss function |
Keyword(4) | Spatial frequency |
Keyword(5) | Color |
1st Author's Name | Naoyuki Ichimura |
1st Author's Affiliation | National Institute of Advanced Industrial Science and Technology(AIST) |
Date | 2018-03-18 |
Paper # | BioX2017-36,PRMU2017-172 |
Volume (vol) | vol.117 |
Number (no) | BioX-513,PRMU-514 |
Page | pp.pp.1-6(BioX), pp.1-6(PRMU), |
#Pages | 6 |
Date of Issue | 2018-03-11 (BioX, PRMU) |