| 摘 要: 基于深度学习的图像压缩感知重构在当前得到了广泛关注。为了进一步提高图像压缩感知的重构质量,提出了一种基于跨阶段记忆增强的图像压缩感知重构网络(Cross-stage Memory Enhancement-based Network,CMENet)。该网络基于迭代软阈值算法进行深度展开,首先引入跨阶段记忆机制以增强阶段之间的特征传递能力和当前阶段特征表达能力,并且利用混洗注意力机制从通道维度和空间维度对特征进行自适应融合。实验结果表明,与现有方法相比,所提出的方法在Set11和BSD68数据集<sup>上均取得了较好的重构性能。</sup>所提出的基于跨阶段记忆增强的图像压缩感知重构网络不仅能获得更高的图像重构质量,而且具有更少的图像重构时间,在效率上也具有优势。 |
| 关键词: 深度学习 图像压缩感知 混洗注意力机制 跨阶段特征增强 |
|
中图分类号: TP391
文献标识码:
|
| 基金项目: 国家自然科学基金（62471241） |
|
| Cross-stage Memory Enhancement-based Network for Image Compressed Sensing Reconstruction |
|
Yang Anqi, Zhang Dengyin
|
School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing Jiangsu 210003
|
| Abstract: Deep learning-based image compressive sensing reconstruction has attracted extensive attention in recent years. To further improve reconstruction quality, this paper proposes a Cross-stage Memory Enhancement-based Network, termed CMENet. The proposed network is built on a deep unfolding framework inspired by the iterative shrinkage-thresholding algorithm. Specifically, a cross-stage memory enhancement mechanism is introduced to strengthen inter-stage feature transmission and improve the representation ability of stage-wise features. In addition, a shuffle attention module is embedded into the reconstruction stage to adaptively recalibrate features from both channel and spatial dimensions. Experimental results on the Set11 and BSD68 datasets demonstrate that the proposed CMENet achieves better reconstruction performance than several existing methods. Moreover, CMENet maintains competitive reconstruction efficiency while improving reconstruction quality, indicating a favorable balance between performance and computational cost. |
| Keywords: Deep learning Image compressive sensing Shuffle attention mechanism Cross-stage feature enhancement |
