Adaptive Compressive Sensing Image Recovery Using Block-Variance-Based Sampling

The paper proposes an adaptive sampling and reconstruction procedure for image denoising, exploiting the variance of image blocks to distinguish between flat and busy regions. This classification is used to make a decision on the number of samples to discard from each block: more samples can be discarded in flat regions compared to busy ones. This optimizes the denoising process while preserving block quality. Additionally, an image fusion technique is applied to enhance the quality of the reconstructed/denoised image. The procedure is iteratively repeated. It is tested on several images corrupted by Gaussian noise. The proposed approach is compared with a conventional non-adaptive sampling for image denoising and reconstruction and a $3 \times 3$ mean filtering approach. The results show PSNR improvement for all considered cases.