Space-variant Neural NetworkApproach toBlind ImageDeconvolution
Tanweer Ahmad Cheema, Ijaz Mansoor Qureshi, Abdul Jalil · 2006
A newspace-variant neural network basedblind image deconvolution method isproposed forrestoring degraded images and thespace-variant blursimultaneously. Theneural network modelis basedonanautoregressive movingaverage (ARMA)process in whichtheAR partoftheneural network defines theimagemodel anditscoefficients, whiletheMA partdefines thedegradation process. Since thebluraffects textured region andsmooth regions differently theimageisdivided intoblocks, whichhavebeen categorized into fourclasses according totheir activity. Furthermore itisassumed thatblurisspace-invariant ineachblock.This classification enhances thetexture ofrelatively textured regions while suppressing thenoise inrelatively smoother backgrounds. A newcost function motivated bythehumanvisual perception system isalso proposed fortheneural network. Thiscomprises oftwonew termsformatching thesecond order local statistics ofeachblock in ordertoimprove thevisual quality oftherestored image. Improvement insignal tonoise ratio (ISNR) andnormalized mean square error (NMSE)oftheestimated blur havebeenusedasfigures ofmerit. Ourproposed algorithm hasgiven better results interms of ISNRandNMSE compared tosomeoftheresults available inthe