Automated Classification of Watermelon Quality Using Non-flicking Reduction and HMM Sequences Derived from Flicking Sound Characteristics

It is challenging for buyers around the globe to identify good quality fruit. For several kinds of fruit, it may be difficult for buyers to determine the fruit quality by appearance. The ability to select only good quality fruit without cutting or cleaving is useful because buyers will not waste money ordering undesirable fruit. To decrease the chances of buyers purchasing sub-standard watermelons, a method that automatically classifies watermelon quality using flicking sounds is proposed. First, preprocessing was used to reduce the non-flicking parts of the signals. Then, Mel Frequency Cepstral Coefficients (MFCCs) and delta accelerator coefficients were extracted from the flicking signals. Finally, the extracted features were recognized by sequences of Hidden Markov acoustic models derived from the characteristics of good and bad flicking sounds. In the experiments, average quality classification rates of 95.0%, 97.0%, 98.0%, 98.0% and 98.0% are obtained by using one to five flicks, respectively. The average computation time spent on one to five flicks is 31.25, 46.02, 63.07, 79.92 and 98.74 milliseconds, respectively. Based on the obtained classification rates and the computation time, the results indicate that the proposed automated method is very efficient and even better at determining watermelon quality than humans are.