Block-based compressive imaging with a swin transformer

Block-based compressive imaging (BCI) is based on the compressive sensing principle, which uses a spatial light modulator and a low-resolution detector to perform parallel high-speed sampling, followed by super-resolution algorithm to reconstruct target image. When compared with traditional compressive imaging, BCI reduces the computational effort but introduces block artifacts. This paper proposes a data-driven deep neural network based on the swin transformer called SwinBCI, which introduces the local attention and shifted window mechanisms to improve the target image reconstruction quality. By using the dataset to train the model to obtain priori knowledge and performing graphics processing unit-accelerated computation, the computation time is greatly reduced to realize real-time BCI. We achieve better reconstruction performances with cake cutting-Hadamard matrix sampling than with Bernoulli matrix sampling. Comparison with three other classical compressed sensing reconstruction methods on four common image datasets and images acquired experimentally using the actual BCI system show that SwinBCI achieves faster high-quality reconstruction at each sampling rate.