A Design and Development of Good and Bad Cocoon Classification Models using Convolutional Neural Network and Transfer Learning

This research aims to design and development of good and bad cocoon classification models using convolutional neural network and transfer learning. There are 3 steps: 1) data acquisition, 2) data preparation, and 3) model development. The data used was a WAV audio file, recorded using a microphone and computer program, which had experts shake out 500 good cocoons and 500 bad cocoons, for a total of 1000 files. The data was converted from audio data into images using Mel spectrogram technique, The data was converted from audio data into images using the Mel Spectrum technique into 1000 images and crop the image using 2 methods, including 1) finding the center of the image and 2) resizing the image. Using a convolutional neural network. and transfer learning in modeling. The data is divided into two parts, 800 images of training data and 200 images of test data. The results show that the convolutional neural network model has an accuracy of 100 percent, Since the images obtained after cropping the images from both methods, there are still prominent features in the center of the image and the images used for training are images of the spectrum of good and bad cocoons, reducing noise or bias in classification model. As for the MobileNetV2 transfer learning model has an accuracy of 95 percent, and the NASNetMobile has an accuracy of 94.5 percent.