Ultra-lightweight Image Compressive Sensing Reconstruction Algorithm Based on Knowledge Distillation

Abstract

Deep neural networks have been shown to improve the quality of image compressive sensing reconstruction, but they are often limited in practical applications due to computational complexity. To address this issue, this paper proposes an ultra-lightweight image compressive sensing reconstruction network. In this network, an adaptive bipolar sampling module is used for information extraction, while sub-pixel convolution and depth-separable convolution are employed for reconstruction to reduce network parameters. Additionally, an improved knowledge distillation algorithm is used to train the network, which further enhances its reconstruction performance. Experimental results show that the proposed ultra-lightweight network has the lowesr computational complexity and the faster reconstruction speed.