1 C Lassification of F Ood K Ernels with I Mpact a Coustics T Ime - 1 F

The authors Nuri F. Ince, Post-Doctoral Associate and Ahmed H. Tewfik , Professor, are with Department of Electrical and 5 Computer Engineering, University of Minnesota, Minneapolis, MN, USA. Ibrahim Onaran, PhD Student, and A. Enis Cetin, 6 Professor, are with Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey. Habil Kalkan, PhD 7 Student, and Yasemin Yardimci, Professor, are with Informatics Institute, Middle East Technical University, Ankara, Turkey. Tom 8 Pearson, ASABE Member, Engineer, is with United States Department of Agriculture Agricultural Research Service, Manhattan, KS, 9 USA. Corresponding author: Nuri F. Ince, Department of Electrical and Computer Engineering, 4-147C EE/CSCI Building, 200 10 Union St. SE, Minneapolis, MN 55455; phone: 612-625-5006; e-mail: [email protected]. 11 12 Abstract. A new adaptive time-frequency (t-f) analysis and classification procedure is applied to impact acoustic 13 signals for detecting hazelnuts with cracked shells and damaged wheat kernels. The impact acoustic signals were 14 recorded as kernels dropped onto a steel plate. The resulting signals were segmented with a flexible local discriminant 15 bases (F-LDB) procedure. The F-LDB method combines local cosine packet analysis and a frequency axis clustering 16 approach which supports individual time and frequency band adaptation. Discriminant features were extracted from the 17 adaptively segmented acoustic signal, sorted according to a Fisher class separability criterion, post processed by 18 principal component analysis and fed to linear discriminant classifier. Experimental results establish the superior 19 performance of the proposed approach when compared to prior techniques reported in the literature or used in the field. 20 The new approach achieved classification accuracies in paired separation of undamaged wheat from insect damaged 21 kernels with exit holes, kernels with hidden damage at the pupal stage and scab damaged kernels with 96%, 82% and 22 94%, respectively. For the cracked undamaged shell hazelnut separation, the overall accuracy was 97.1%. 23