DRCT: Saving Image Super-Resolution away from Information Bottleneck

DRCT: Saving Image Super-Resolution away from Information Bottleneck

15 Apr 2024 | Chih-Chung Hsu, Chia-Ming Lee, Yi-Shiuan Chou
The paper "DRCT: Saving Image Super-Resolution away from Information Bottleneck" by Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou addresses the issue of information bottleneck in Vision Transformer-based models used for image super-resolution (SISR). The authors observe that as the network depth increases, the intensity of feature maps decreases sharply, leading to a loss of spatial information and potentially limiting the model's performance. To mitigate this, they propose the Dense-residual-connected Transformer (DRCT), which incorporates dense-residual connections within residual blocks to stabilize the information flow and enhance the receptive field. This approach allows the model to capture long-range dependencies more effectively without increasing computational complexity. Experimental results on benchmark datasets and the NTIRE-2024 Image Super-Resolution (x4) Challenge demonstrate that DRCT outperforms state-of-the-art methods, achieving superior performance with simpler model architectures. The paper also discusses related works, problem statements, and the methodology behind DRCT, including its architecture and training strategies.The paper "DRCT: Saving Image Super-Resolution away from Information Bottleneck" by Chih-Chung Hsu, Chia-Ming Lee, and Yi-Shiuan Chou addresses the issue of information bottleneck in Vision Transformer-based models used for image super-resolution (SISR). The authors observe that as the network depth increases, the intensity of feature maps decreases sharply, leading to a loss of spatial information and potentially limiting the model's performance. To mitigate this, they propose the Dense-residual-connected Transformer (DRCT), which incorporates dense-residual connections within residual blocks to stabilize the information flow and enhance the receptive field. This approach allows the model to capture long-range dependencies more effectively without increasing computational complexity. Experimental results on benchmark datasets and the NTIRE-2024 Image Super-Resolution (x4) Challenge demonstrate that DRCT outperforms state-of-the-art methods, achieving superior performance with simpler model architectures. The paper also discusses related works, problem statements, and the methodology behind DRCT, including its architecture and training strategies.
Reach us at info@study.space