Research On Video Snapshot Compressive Imaging Algorithm Based On Provable Bounded Denoisers

Snapshot compressive imaging(SCI)is an important technology in the field of computational imaging,which can effectively capture high-dimensional signals.As a key component of the SCI system,the reconstruction algorithms for SCI have been greatly developed in recent years due to the influence of deep learning.The plug-and-play(PnP)SCI algorithms that leverage deep denoisers have achieved excellent performance,and their convergences have been guaranteed under the assumption of bounded denoisers.However,it is difficult to explicitly prove the bounded properties of existing deep denoisers due to complex network architectures.To address these issues,this paper studies provable bounded denoisers based on the tight frame learned by supervised learning.These denoisers are integrated into the PnP framework to construct video SCI algorithms with provable convergence.The specific research contents are as follows:Firstly,to address the issue of low performance of denoisers based on data-driven tight frames,the supervised single tight frame is extended to dual tight frames,and a provable bounded denoiser based on dual tight frames is proposed.The threshold plays a crucial role in the denoiser based on tight frame,and we set it proportional to the noise level,with the proportionality coefficient adaptively obtained from the input instance by the coefficient generating network(CNet)to acquire a spatially varying threshold map.Then the dual tight frames-based denoiser is plugged into the PnP framework,resulting in a video SCI algorithm,and its convergence is proven.Experimental results demonstrate the effectiveness of the proposed denoiser and SCI algorithm.Secondly,to further improve the performance of tight frame-based denoisers and SCI algorithms,the dual tight frames are extended to the double tight frames to enhance their representation ability,and then the Swin-Transformer block is incorporated into the CNet to capture global information.A provable bounded deep denoiser is designed by using the double tight frames and the CNet which can capture global information.Subsequently,this bounded denoiser is plugged into the PnP framework to propose a video SCI algorithm,and its fixed-point convergence is proven.Experimental results demonstrate that the dual tight frames-based denoiser outperforms the double tight frames-based denoiser in denoising performance.Furthermore,the proposed algorithm achieves performances comparable to the benchmark algorithms in video SCI reconstruction tasks.Finally,to address the problem that the proposed denoisers and algorithms ignored the redundant information in the temporal domain of video sequences,3D convolution is incorporated into the CNet to fuse information in both temporal and spatial domains simultaneously.Based on this,a video denoiser is designed by using the double tight frames and the CNet fused with spatiotemporal information.The denoiser takes continuous image sequences as input,which can fully consider temporal correlation,and its bounded property can also be explicitly proved.Integrating this video denoiser into the PnP framework and constructing a convergent video SCI algorithm.Experimental results show that the proposed algorithm can achieve better reconstruction performances than the state-of-the-art PnP-based SCI algorithm.