32
3
Vol.32,No.3May,2006
2006
5
ACTAAUTOMATICASINICA
1)
(
100083)
(E-mail:xuke@ustb.edu.cn)
128×128Hu
LVQ
81.5%.
TP751.1
FeatureExtractionBasedonAmplitudeSpectrumandMoment
InvariantsandItsApplication
XUKe
LIWen-Feng
YANGChao-Lin
(NationalEngineeringResearchCenterofAdvancedRollingTechnology,UniversityofScienceand
Technology,Beijing100083)
(E-mail:xuke@ustb.edu.cn)
AbstractPrinciplesandcharactersofFouriertransformandmomentinvariantswerestud-
ied.Featureextractionbasedonamplitudespectrumandmomentinvariantswasproposedandappliedtoclassificationofsurfacedefectsofmediumandheavyplates.Surfaceimagescapturedfromaproductionlineofheavyandmediumplatesweredividedintosub-imagesofsize128×128.Amplitudespectrumsofsub-imageswereobtainedwithfastFouriertrans-form.HuMomentInvariantsofamplitudespectrumswerecalculatedasfeatures,whichwerenotonlytranslationinvariantandrotationinvariantbutalsoinsensitivetonoisesandunevenillumination.ThefeatureswerefedintoclassifiersbasedonLVQnetworks.Resultsoftrainingandtestingwithsamplesofdefectsshowedthatthefeatureswereapplicabletoclassficationofsurfacedefectsofmediumandheavyplates,andtheclassificationratewas81.5%.Keywords
Amplitudespectrum,momentinvariant,featureextraction,surfaceinspection
3
471
472
32
3
473
Fig.4(a)Fig.4(b)Fig.4(c)
0.0210.0210.021
0.0510.0510.051
0.1060.1060.106
0.1020.1020.103
0.2080.2070.209
0.1310.1310.133
0.2070.2070.211
4.4
487
292
195
LVQ
2
3
3
80%.
2
Table2
Resultsoftraining
(%)
45451913133131195
39361310102427159
86.780.068.476.976.978.1987.181.5
474
32