GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention
Most GAN-based methods utilize semantic layouts as input for generating realistic images. However, these layouts primarily consist of object contours and often lack detailed information, leading to suboptimal image quality in the generated outputs. To address this limitation, we propose a novel GAN...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10767714/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1850126410998349824 |
|---|---|
| author | Xinhua Dong Chuang Li Zhigang Xu Hongmu Han Lifeng Jiang |
| author_facet | Xinhua Dong Chuang Li Zhigang Xu Hongmu Han Lifeng Jiang |
| author_sort | Xinhua Dong |
| collection | DOAJ |
| description | Most GAN-based methods utilize semantic layouts as input for generating realistic images. However, these layouts primarily consist of object contours and often lack detailed information, leading to suboptimal image quality in the generated outputs. To address this limitation, we propose a novel GAN architecture called LMCGAN designed specifically for synthesizing high-quality images. LMCGAN introduces a generator network structured around the laplacian pyramid, enabling the simultaneous generation of multi-scale feature maps.This approach allows the model to capture finer details at different resolutions, enhancing the overall realism of the generated images.To further improve the utilization of semantic maps, we integrate a multi-scale channel attention (MSCA) mechanism.This mechanism effectively focuses on channel-specific information in complex scenes, which is crucial for preserving essential details that may otherwise be lost. During the feature fusion phase, we implement a feature fusion block (FFBL) that is designed to capture important relationships across various scales. This block facilitates the integration of information from different resolutions, ensuring that the final output retains critical features. Additionally, we adopt a combination of conditional and unconditional methods to reduce noise during the training process, leading to more stable and effective training dynamics. Extensive experiments conducted on challenging datasets demonstrate that LMCGAN significantly outperforms existing methods in terms of both visual quality and quantitative evaluation metrics. The results indicate that our architecture not only generates more realistic images but also excels in preserving intricate details, marking a substantial advancement in the field of image synthesis using GANs. |
| format | Article |
| id | doaj-art-c13e78c870434966a55d65dceda7cf50 |
| institution | OA Journals |
| issn | 2169-3536 |
| language | English |
| publishDate | 2024-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Access |
| spelling | doaj-art-c13e78c870434966a55d65dceda7cf502025-08-20T02:33:55ZengIEEEIEEE Access2169-35362024-01-011217801017802110.1109/ACCESS.2024.350657710767714GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel AttentionXinhua Dong0https://orcid.org/0009-0008-2491-1593Chuang Li1https://orcid.org/0009-0000-0263-2173Zhigang Xu2https://orcid.org/0000-0003-4007-4779Hongmu Han3https://orcid.org/0000-0002-6909-5242Lifeng Jiang4https://orcid.org/0009-0007-7917-8168School of Computer Science, Hubei University of Technology, Wuhan, ChinaSchool of Computer Science, Hubei University of Technology, Wuhan, ChinaSchool of Computer Science, Hubei University of Technology, Wuhan, ChinaSchool of Computer Science, Hubei University of Technology, Wuhan, ChinaSchool of Computer Science, Hubei University of Technology, Wuhan, ChinaMost GAN-based methods utilize semantic layouts as input for generating realistic images. However, these layouts primarily consist of object contours and often lack detailed information, leading to suboptimal image quality in the generated outputs. To address this limitation, we propose a novel GAN architecture called LMCGAN designed specifically for synthesizing high-quality images. LMCGAN introduces a generator network structured around the laplacian pyramid, enabling the simultaneous generation of multi-scale feature maps.This approach allows the model to capture finer details at different resolutions, enhancing the overall realism of the generated images.To further improve the utilization of semantic maps, we integrate a multi-scale channel attention (MSCA) mechanism.This mechanism effectively focuses on channel-specific information in complex scenes, which is crucial for preserving essential details that may otherwise be lost. During the feature fusion phase, we implement a feature fusion block (FFBL) that is designed to capture important relationships across various scales. This block facilitates the integration of information from different resolutions, ensuring that the final output retains critical features. Additionally, we adopt a combination of conditional and unconditional methods to reduce noise during the training process, leading to more stable and effective training dynamics. Extensive experiments conducted on challenging datasets demonstrate that LMCGAN significantly outperforms existing methods in terms of both visual quality and quantitative evaluation metrics. The results indicate that our architecture not only generates more realistic images but also excels in preserving intricate details, marking a substantial advancement in the field of image synthesis using GANs.https://ieeexplore.ieee.org/document/10767714/Channel attentionfeature fusionGANLaplacian image pyramidsemantic image synthesis |
| spellingShingle | Xinhua Dong Chuang Li Zhigang Xu Hongmu Han Lifeng Jiang GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention IEEE Access Channel attention feature fusion GAN Laplacian image pyramid semantic image synthesis |
| title | GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention |
| title_full | GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention |
| title_fullStr | GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention |
| title_full_unstemmed | GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention |
| title_short | GAN for Semantic Image Synthesis With Laplacian Pyramid and Multi-Scale Channel Attention |
| title_sort | gan for semantic image synthesis with laplacian pyramid and multi scale channel attention |
| topic | Channel attention feature fusion GAN Laplacian image pyramid semantic image synthesis |
| url | https://ieeexplore.ieee.org/document/10767714/ |
| work_keys_str_mv | AT xinhuadong ganforsemanticimagesynthesiswithlaplacianpyramidandmultiscalechannelattention AT chuangli ganforsemanticimagesynthesiswithlaplacianpyramidandmultiscalechannelattention AT zhigangxu ganforsemanticimagesynthesiswithlaplacianpyramidandmultiscalechannelattention AT hongmuhan ganforsemanticimagesynthesiswithlaplacianpyramidandmultiscalechannelattention AT lifengjiang ganforsemanticimagesynthesiswithlaplacianpyramidandmultiscalechannelattention |