Kronecker reparameterized large kernel for image compressed sensing

Image compressed sensing methods utilizing deep neural networks have achieved remarkable performance improvements compared with traditional ones, regarding both reconstruction quality and efficiency. However, the factors that affect the reconstruction quality of such deep compressed sensing methods remain unclear. In this paper, we reveal that one important factor is the size of the Effective Receptive Field (ERF), based on which we propose a novel Convolutional Compressed Sensing Network with the Kronecker Reparameterized Large Kernel (KR-CCSNet). Specifically, to enlarge the ERF and achieve superior reconstruction quality, we propose the Kronecker Reparameterized Large Kernel Sampling Network (KR-LKSN) for the sampling phase. KR-LKSN not only delivers better reconstruction quality, reduced computation, and fewer parameters, but also shows great potential for deployment on resource-constrained edge sensors, owing to the lightweight design of its sampling module. For the reconstruction network, we design an Adaptive Reconstruction Module (ARM) to leverage multi-scale information from measurements via gated attention, which further enlarges the ERF during the reconstruction phase to generate high-quality images. Extensive experiments demonstrate the effectiveness of KR-CCSNet on Set5, Set14, and BSDS100. For instance, our method outperforms MR-CCSNet by an average PSNR of 0.35 dB on Set5 and Set14 across six compression ratios. Our source codes are released at https://github.com/Will0x6c5f/KRCCSNet.