Presentation | 2021-03-04 Learning Convolutional Neural Networks with Spatial Frequency Loss Naoyuki Ichimura, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | The pixel-wise L2 and pixel-wise L1 losses have been commonly used to measure the consistency between images in learning convolutional neural networks~(CNNs) for image generation tasks. However, using the losses poses the well-known problem of producing blurry images. This paper presents a learning method using a novel loss called the spatial frequency loss~(SFL) to mitigate the problem. The blurs in generated images show the lack of high spatial frequency components and the degree of deficiency depends on the frequency response of CNNs. In order to analyze the frequency response of CNNs, a Laplacian filter bank that has a band-pass property is added to CNNs to extract features in subbands of images. Then the SFL is defined by the sum of the L2 losses of the features over subbands and the losses corresponding to high spatial frequency components are emphasized by weighting in learning. Experimental results for image inpainting using CNNs demonstrate that learning with the SFL is fairly useful to reduce the blurs and produce the fine texture details in generated images. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Convolutional neural networks / Loss function / Spatial frequency / Image inpainting |
Paper # | PRMU2020-73 |
Date of Issue | 2021-02-25 (PRMU) |
Conference Information | |
Committee | PRMU / IPSJ-CVIM |
---|---|
Conference Date | 2021/3/4(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Online |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Computer Vision and Pattern Recognition for specific environment |
Chair | Yoichi Sato(Univ. of Tokyo) |
Vice Chair | Akisato Kimura(NTT) / Masakazu Iwamura(Osaka Pref. Univ.) |
Secretary | Akisato Kimura(Mobility Technologies) / Masakazu Iwamura(Chubu Univ.) |
Assistant | Takashi Shibata(NTT) / Masashi Nishiyama(Tottori Univ.) |
Paper Information | |
Registration To | Technical Committee on Pattern Recognition and Media Understanding / Special Interest Group on Computer Vision and Image Media |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Learning Convolutional Neural Networks with Spatial Frequency Loss |
Sub Title (in English) | |
Keyword(1) | Convolutional neural networks |
Keyword(2) | Loss function |
Keyword(3) | Spatial frequency |
Keyword(4) | Image inpainting |
1st Author's Name | Naoyuki Ichimura |
1st Author's Affiliation | National Institute of Advanced Industrial Science and Technology(AIST) |
Date | 2021-03-04 |
Paper # | PRMU2020-73 |
Volume (vol) | vol.120 |
Number (no) | PRMU-409 |
Page | pp.pp.25-30(PRMU), |
#Pages | 6 |
Date of Issue | 2021-02-25 (PRMU) |