wganBCS: Block-wise image compressive sensing and reconstruction model using adversarial training to eliminate block effects

• Item type: SIG Technical Reports(1)
• 公開日: 2021-11-18
• タイトル: wganBCS: Block-wise image compressive sensing and reconstruction model using adversarial training to eliminate block effects
• タイトル / 言語: en
• タイトル: wganBCS: Block-wise image compressive sensing and reconstruction model using adversarial training to eliminate block effects
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Graduate School of Computer and Information Sciences, Hosei University／School of Software, Northwestern Polytechnical University
• 著者所属: Graduate School of Computer and Information Sciences, Hosei University
• 著者所属(英): en
• 著者所属(英): Graduate School of Computer and Information Sciences, Hosei University / School of Software, Northwestern Polytechnical University
• 著者所属(英): en
• 著者所属(英): Graduate School of Computer and Information Sciences, Hosei University
• 著者名: Boyan, Chen; Kaoru, Uchida
• 著者名(英): Boyan, Chen; Kaoru, Uchida
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: The famous block-wise compressive sensing (BCS) paradigm can greatly reduce the memory consumption of sensing matrix compared with traditional image compressive methods. However, BCS paradigm still suffers from two issues. One is that block-wise sensing causes heavy block effect on the reconstructed image, which leads to degradation in the image quality metrics. Another is that the sate of art block wise image compressive sensing methods only use mean square error loss function to optimize their models, which causes the reconstructed images over smoothed. In this paper, we incorporate generative adversarial training into BCS paradigm and propose a new block wise image compressive sensing and reconstruction model called wganBCS, which a combination of traditional L2 loss and the wasserstein loss are used to optimize the model. We propose a modified wasserstein GAN (WGAN) network to deal with the block effect caused by the block wise compressive sensing. Specifically speaking, the generator network will minimize the wasserstein distance calculated by the critic network to keep the reconstructed images visually authentic to ground truth images. Experimental result shows that our model is superior both in visual authenticity and the image quality metrics compared to most state of art image compressive sensing methods.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: The famous block-wise compressive sensing (BCS) paradigm can greatly reduce the memory consumption of sensing matrix compared with traditional image compressive methods. However, BCS paradigm still suffers from two issues. One is that block-wise sensing causes heavy block effect on the reconstructed image, which leads to degradation in the image quality metrics. Another is that the sate of art block wise image compressive sensing methods only use mean square error loss function to optimize their models, which causes the reconstructed images over smoothed. In this paper, we incorporate generative adversarial training into BCS paradigm and propose a new block wise image compressive sensing and reconstruction model called wganBCS, which a combination of traditional L2 loss and the wasserstein loss are used to optimize the model. We propose a modified wasserstein GAN (WGAN) network to deal with the block effect caused by the block wise compressive sensing. Specifically speaking, the generator network will minimize the wasserstein distance calculated by the critic network to keep the reconstructed images visually authentic to ground truth images. Experimental result shows that our model is superior both in visual authenticity and the image quality metrics compared to most state of art image compressive sensing methods.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN10438399
• 書誌情報: 研究報告オーディオビジュアル複合情報処理（AVM）; 巻 2021-AVM-115, 号 2, p. 1-6, 発行日 2021-11-18
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8582
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会