Block-DCT compressed capsule endoscopic images generally suffer annoying blocking artifacts. Many approaches have been proposed to deblock with tissue-pattern retention. However, their performances of self-adaptively filtering are still weak, and no feasible method has been defined yet to precisely distinguish between real edges and mosaic effects. In this paper, a new metric developed from the mean squared difference of slope (MSDS) is introduced to achieve more accurate block evaluation, then a novel deblocking method called guided frequency filter (GFF) is proposed. GFF is a frequency domain filter adopted from guided filter theory. By utilizing human visual system (HVS) and the new metric, GFF achieves smoothness-degree-adaptive and frequency adaptive. Experimental results show that this method yields to not only good image visual quality with edge-preservation but also an average PSNR improvement.