Uniform block compressed sensing cannot separate the important region from the background for image signals effectively. A new notion of saliency-based adaptive block compressive sampling method is proposed. According to the saliency of the image signal

the quadtree algorithm is introduced to separate the important block and background block adaptively. The amount of observation samples is assigned dynamically to improve the quality of image reconstruction in salient regions. A high sampling rate is set for the important regions

while a low value is used for the background regions. Experimental results validate its rationality and effectiveness. Compared with uniform block compressed sensing

the proposed method needs fewer observations

and has a better performance in PSNR and MSSIM with a shorter running time.