Durian Ripeness Classification from the Knocking Sounds Using Convolutional Neural Network
Weangchai Kharamat, Manop Wongsaisuwan, Norrarat Wattanamongkhol
Abstract
Durian is one of the most popular fruits in Thailand. The ripeness checking of durian is very important for fruit marketing and durian exports. To check the ripeness of durians in Thailand, there are many existing methods to increase the efficiency of the durian harvesting and exporting. A common method is knocking a few times using a rubber-tipped stick. This paper proposes a method to indicate the ripeness of durians. The total of 30 durian samples is used to collect the knocking sound data. Then we separate sound data into 0.3 seconds of each durian knocking sound interval. The dataset was recorded from a durian exporter in the Chanthaburi Province using a smartphone. Our method applies the convolutional neural network (CNN) to classify the durian sound. The durian sound received from the percussion is then contained in a dataset. The dataset. We, then, classify each dataset into three classes of ripeness: ripe, mid-ripe, and unripe. A process of audio extraction applies Mel-frequency cepstral coefficient spectrogram (MFCC) to extraction a feature of durian sound and utilizes a set of feature as training data. This dataset is used as the training information for CNN. An experimental results of our model shows an accuracy around 90.78% of validation data and 89.47% of testing data. Source code to train the model is available at https://github.com/KENGKKUN/Durian-ripeness-sound-regnization.git.