Non-reference quality assessment of image processing techniques in infrared non-destructive testing.

aa common procedure to compare the ability of Infrared for NDT (IRNDT) image processing techniques for image enhancement involves experts who visually compares the outcomes of the techniques and determines which technique, for some given parameters, works better by counting the number of defects made visible in a particular image [Zalamedaetal.,2003].Theresultsofthesecomparisonsbecomeargumentstoexplainwhyanewtechniqueismoreconvenientthanprevioustechniquesinagiveninspection. On theother hand,SNR(Signalto NoiseRatio)ordetectivityisaquantitativemethod forcomparison [Pickeringand almond,2008], [Omarand Zhou,2008]. Thisquantitative comparisonisconstrainedbythedefinitionofareferencenon-defectivearea.Inawidesense,itslocationisnotpreciselyidentifiedsinceitmaynotbeknowninadvancewhere thedefectsare,ifpresentatall[Benitezetal.,2007].SNRdefinitionisusefultomeasurethedetectivityofonedefectintheimagebutitdoesnotprovideanassessmentofthe overallimagecontrastenhancementprovidedbythespecifictechnique.Recently[Florez-OspinaandBenitez,2014]proposed aSNRversionthatpermitstheglobalevaluation of defect visibility. This work also presented the image quality assessment (IQa) index IRIQI (Infrared Image Quality), that makes use ofthe structural information and spatial dependencyof pixelsin images processedbyIRNDT[Wang etal., 2004],[Wang andBovik, 2002].However, theseapproaches stilldependon apriori knowledge ofdefective andnon-defective regions locationintheimage. InIRNDTitisimportant tohaveNo-Reference(NR)IQametricssinceoriginal orpristineimagesarenotavailableanddefects locationisunknown. Inparticular, DIVIINE[MoorthyandBovik,2011]haveshowntocorrelatehighlywithhumansubjectivescores.Thismethodattemptstomodelthe’natural’ statistics found in pristine images [Moorthy and Bovik, 2010]. DIVIINE IQaof IRNDT show a high correlation between quality predictions and subjective evaluation [RamirezRozo et al., 2013]. Nevertheless, this study selected only a subset of processed images from a sequence for subjective evaluation and it was observed that DIVIINE equally scores images with different visual content (i.e defect visibility). This suggests that ’natural’ features presented in visible spectrum images do not correctly model processed IRNDT images. The limitations previously described indicate that the proper comparison (i.e non-referenced and highly correlated with human subjective scores) of image processingtechniquesininfrarednon-destructivetesting(IRNDT)remainsdifficult.Thissituationcreatesthequestions:HowtodefineappropriatefeaturesinIRNDTimagesthat representdefectvisibilitydegradationinaprocessedimagessequence?HowtodefineandvalidateaNRIQametrictopredicthumanvisualassessmentofIRNDTimages?The main goal of this project is the developing and validating of NR-IQa metrics that represent defect visibility and predict human visual evaluation of IRNDT images