With original image $o$ fixed, the higher, the better. If PSRN($o,d$) is bigger than PSRNPSNR($o,n$), you can claim an improvement in the decibel measure. It is difficult to interpret PSRN($d,n$), sinceyousince you compare two more-or-less noisy data.
Is PSRNPSNR a good measure of subjective image quality remains an open question.