The ability of snapshot compressive imaging (SCI) systems to efficiently capture high-dimensional (HD) data has led to an inverse problem, which consists of recovering the HD signal from the compressed and noisy measurement. While reconstruction algorithms grow fast to solve it with the recent advances of deep learning, the fundamental issue of accurate and stable recovery remains. To this end, we propose deep equilibrium models (DEQ) for video SCI, fusing data-driven regularization and stable convergence in a theoretically sound manner. Each equilibrium model implicitly learns a nonexpansive operator and analytically computes the fixed point, thus enabling unlimited iterative steps and infinite network depth with only a constant memory requirement in training and testing. Specifically, we demonstrate how DEQ can be applied to two existing models for video SCI reconstruction: recurrent neural networks (RNN) and Plug-and-Play (PnP) algorithms. On a variety of datasets and real data, both quantitative and qualitative evaluations of our results demonstrate the effectiveness and stability of our proposed method. The code and models are available at: https://github.com/IndigoPurple/DEQSCI.
In order to capture the spatial-spectral information of the scene, various techniques have been proposed. Different from the widely used scanning-based methods, spectral snapshot compressive imaging (SCI) utilizes the idea of compressive sensing to compressively capture the 3D spatial-spectral data-cube in a single-shot 2D measurement and thus it is efficient, enjoying the advantages of high-speed and low bandwidth. However, the reconstruction process, i.e., to retrieve the 3D cube from the 2D measurement, is an ill-posed problem and it is challenging to reconstruct high quality images. Previous works usually use 2D convolutions and preliminary attention to address this challenge. However, these networks and attention do not exactly extract spectral features. On the other hand, 3D convolutions can extract more features in a 3D cube, but increase computational cost significantly. To balance this trade-off, in this paper, we propose a hybrid multi-dimensional attention U-Net (HMDAU-Net) to reconstruct hyperspectral images from the 2D measurement in an end-to-end manner. HMDAU-Net integrates 3D and 2D convolutions in an encoder–decoder structure to fully utilize the abundant spectral information of hyperspectral images with a trade-off between performance and computational cost. Furthermore, attention gates are employed to highlight salient features and suppress the noise carried by the skip connections. Our proposed HMDAU-Net achieves superior performance over previous state-of-the-art reconstruction algorithms.
3 December 2025 in IEEE Transactions on Multimedia
Digital cameras consume $\sim 0.1$ microjoule per pixel to capture and encode video, resulting in a power usage of $\sim 20$W for a 4K sensor operating at 30 fps. Imagining gigapixel cameras operating at 100-1000 fps, the current processing model is unsustainable. To address this, physical layer compressive measurement has been proposed to reduce power consumption per pixel by 10-100×. Video Snapshot Compressive Imaging (SCI) introduces high frequency modulation in the optical sensor layer to increase effective frame rate. A commonly used sampling strategy of video SCI is Random Sampling (RS) where each mask element value is randomly set to be 0 or 1. Similarly, image inpainting (I2P) has demonstrated that images can be recovered from a fraction of the image pixels. Inspired by I2P, we propose Ultra-Sparse Sampling (USS) regime, where at each spatial location, only one sub-frame is set to 1 and all others are set to 0. We then build a Digital Micro-mirror Device (DMD) encoding system to verify the effectiveness of our USS strategy. Ideally, we can decompose the USS measurement into sub-measurements for which we can utilize I2P algorithms to recover high-speed frames. However, due to the mismatch between the DMD and CCD, the USS measurement cannot be perfectly decomposed. To this end, we propose BSTFormer, a sparse TransFormer that utilizes local Block attention, global Sparse attention, and global Temporal attention to exploit the sparsity of the USS measurement. Extensive results on both simulated and real-world data show that our method significantly outperforms all previous state-of-the-art algorithms. Additionally, an essential advantage of the USS strategy is its higher dynamic range than that of the RS strategy. Finally, from the application perspective, the USS strategy is a good choice to implement a complete video SCI system on chip due to its fixed exposure time. Code is available at https://github.com/mcao92/BSTFormer.
18 November 2025 in IEEE Journal of Selected Topics in Signal Processing
Image compression is an important technique to reduce the requirement of bandwidth, especially for large-scale information transmission and resource-limited platforms, like drones and robotics. In these cases, due to the limit resource, a low-cost encoder is desired. Toward this end, Block Modulation Video Compression (BMVC) encoder is a potential solution for this task. In this paper, we propose BMVC$+$, an enhanced version of BMVC, which promotes both the encoder and decoder of previous design, making it more efficient and experiment-friendly in real-world applications. Rather than separating the large-scale image into non-overlapping partitioned blocks and then summing all blocks together in previous BMVC encoder, the proposed BMVC$+$ encoder treats the large-scale image as a group of low-resolution patches, which can be seen as scanning strategy to compress patch into one pixel, leading to a low resolution compressed measurement. As for reconstruction, the patch sequences are specially extracted and reconstructed by BMVC$+$ decoder, for which we propose a deep unfolding network with a combination of multi-model unrolling Mamba block and channel-wise self-attention block for both local and global feature calibration and long-distance correlation reconstruction. Extensive experiments on the simulation dataset demonstrate the performance of the proposed method, which shows great potential to be applied in real-world system for low-complexity image compression. The source code and pre-trained models of BMVC+ will be released at https://github.com/Gang-Qu/BMVC-plus.
We consider the reconstruction problem of video compressive sensing (VCS) under the deep unfolding/rolling structure. Yet, we aim to build a flexible and concise model using minimum stages. Different from existing deep unfolding networks used for inverse problems, where more stages are used for higher performance but without flexibility to different masks and scales, hereby we show that a 2-stage deep unfolding network can lead to the state-of-the-art (SOTA) results (with a 1.7dB gain in PSNR over the single stage model, RevSCI) in VCS. The proposed method possesses the properties of adaptation to new masks and ready to scale to large data without any additional training thanks to the advantages of deep unfolding. Furthermore, we extend the proposed model for color VCS to perform joint reconstruction and demosaicing. Experimental results demonstrate that our 2-stage model has also achieved SOTA on color VCS reconstruction, leading to a >2.3dB gain in PSNR over the previous SOTA algorithm based on plug-and-play framework, meanwhile speeds up the reconstruction by >17 times. In addition, we have found that our network is also flexible to the mask modulation and scale size for color VCS reconstruction so that a single trained network can be applied to different hardware systems. The code and models will be released to the public.
The ability of snapshot compressive imaging (SCI) systems to efficiently capture high-dimensional (HD) data has led to an inverse problem, which consists of recovering the HD signal from the compressed and noisy measurement. While reconstruction algorithms grow fast to solve it with the recent advances of deep learning, the fundamental issue of accurate and stable recovery remains. To this end, we propose deep equilibrium models (DEQ) for video SCI, fusing data-driven regularization and stable convergence in a theoretically sound manner. Each equilibrium model implicitly learns a nonexpansive operator and analytically computes the fixed point, thus enabling unlimited iterative steps and infinite network depth with only a constant memory requirement in training and testing. Specifically, we demonstrate how DEQ can be applied to two existing models for video SCI reconstruction: recurrent neural networks (RNN) and Plug-and-Play (PnP) algorithms. On a variety of datasets and real data, both quantitative and qualitative evaluations of our results demonstrate the effectiveness and stability of our proposed method. The code and models are available at: https://github.com/IndigoPurple/DEQSCI .