Compressive sampling based on frequency saliency for remote sensing imaging

Abstract

Abstract In saliency-based compressive sampling (CS) for remote sensing image signals, the saliency information of images is used to allocate more sensing resources to salient regions than to non-salient regions. However, the pulsed cosine transform method can generate large errors in the calculation of saliency information because it uses only the signs of the coefficients of the discrete cosine transform for low-resolution images. In addition, the reconstructed images can exhibit blocking effects because blocks are used as the processing units in CS. In this work, we propose a post-transform frequency saliency CS method that utilizes transformed post-wavelet coefficients to calculate the frequency saliency information of images in the post-wavelet domain. Specifically, the wavelet coefficients are treated as the pixels of a block-wise megapixel sensor. Experiments indicate that the proposed method yields better-quality images and outperforms conventional saliency-based methods in three aspects: peak signal-to-noise ratio, mean structural similarity index, and visual information fidelity.