Recognition of Fruit Types from Striking and Flicking Sounds

Abstract

This paper proposes a method to recognize fruits whose quality, including their ripeness, grades, brix values, and flesh characteristics, cannot be determined visually from their skin but from striking and flicking sounds. Four fruit types consisting of durians, watermelons, guavas, and pineapples were studied in this research. In recognition of fruit types, preprocessing removes the non-striking/non-flicking parts from the striking and flicking sounds. Then the sequences of frequency domain acoustic features containing 13 Mel Frequency Cepstral Coefficients (MFCCs) and their 13 first- and 13 second-order derivatives were extracted from striking and flicking sounds. The sequences were used to create the Hidden Markov Models (HMMs). The HMM acoustic models, dictionary, and grammar were incorporated to recognize striking and flicking sounds. When testing the striking and flicking sounds obtained from the fruits used to create the training set but were collected at different times, the recognition accuracy using 1 through 5 strikes/flicks was 98.48%, 98.91%, 99.13%, 98.91%, and 99.57%, respectively. For an unknown test set, of which the sounds obtained from the fruits that were not used to create the training set, the recognition accuracy using 1 through 5 strikes/flicks were 95.23%, 96.82%, 96.82%, 97.05%, and 96.59%, respectively. The results also revealed that the proposed method could accurately distinguish the striking sounds of durians from the flicking sounds of watermelons, guavas, and pineapples.