Learned Measurement Interpolation for Scalable Compressive Sensing

Deep learning-based joint compressive sensing image acquisition and reconstruction has attracted great attention in the field of image compression and computational sensors. Although these methods show excellent results in execution speed and image quality, each compression ratio needs one specific deep-learning model which causes training costs and memory pressure. To solve the above problems, we propose a Measurements Interpolation strategy (MI) that can be placed before a DL-based reconstruction model. MI takes original measurements, which are the compressed data of the input image, and outputs interpolated measurements with an arbitrary compression ratio by interpolating the original measurements. The experimental results show that our proposal can be integrated with the existing DL-based compressive sensing reconstruction works to achieve scalable sampling and reconstruction while keeping the image quality.