Classification of Hazelnut Kernels by Using Impact Acoustic Time-Frequency Patterns

Hazelnuts with damaged or cracked shells are more prone to infection with aflatoxin producing molds (Aspergillus Flavus). These molds can cause cancer. In this study, we introduce a new approach that separates damaged/cracked hazelnut kernels from good ones by using time-frequency features obtained from impact acoustic signals in an offline step. The proposed technique requires no prior knowledge of the relevant time and frequency locations. It adaptively segments impact signals from a training data set in time using local cosine packet analysis and a Kullback-Leibler criterion to assess the discrimination power of different segmentations. In each resulting time segment, the signal is further decomposed into sub-bands using an undecimated wavelet transform. The most discriminative sub-bands are selected according to the Euclidean distance between the cumulative probability distributions of the corresponding sub band coefficients. The most discriminative sub bands are fed into a linear discriminant analysis classifier. In the online classification step, the algorithm simply computes the learned features from the observed signal and feeds them to the LDA classifier. This allows the system to process 71 nuts/sec. The algorithm achieved a classification accuracy of 96.7% with the 24 most discriminative features, a higher rate than those provided with prior methods.