HIT-Net: A high-information-capacity dual-channel transformer network for video snapshot compressive imaging

Video snapshot compressive imaging is a computational imaging technique based on compressed sensing that acquires high-speed frames through a 2D compressed measurement, followed by an algorithmic reconstruction. Current deep learning-based reconstruction methods operate on single-channel measurements and fail to fully capture complex inter-element dependencies, resulting in detail loss and artifacts in high-precision reconstruction tasks. To address these issues, we propose a high-information-capacity dual-channel transformer network. The network first introduces a dual-channel fusion structure that increases effective input information, thereby reducing the equivalent compression ratio and improving reconstruction quality. We further enhance the transformer by replacing conventional positional encodings with a relative Gaussian weighting scheme and introducing an adaptive multi-head structure together with variable cross section residual connections, significantly strengthening the modeling of inter-element dependencies and high-fidelity detail recovery. Experiments on six standard benchmark datasets (Kobe, Runner, Aerial, Crash, Drop, and Traffic) demonstrate that the proposed method markedly outperforms all compared approaches and is the only one that consistently exceeds 30 dB peak signal-to-noise ratio across all datasets.