Classification model of ‘Toraja’ arabica coffee fruit ripeness levels using convolution neural network approach

Abstract

The purpose of this study is to design a CNN deep learning algorithm model that can classify the maturity level of Arabica coffee fruit based on image, the resulting model can be applied to a coffee bean sorting device based on artificial intelligence so that problems that exist in the process of sorting arabica coffee fruit that meets the standards can be avoided, to improve the quality of arabica Toraja coffee products. The research began from the collection of data in the form of raw Arabica coffee image Toraja as many as 4000 images of arabica coffee fruit with 4 categories, half-cooked, perfectly ripe, and mature old. CNN basic architecture is created using images with a size of 128x128 pixels, 4 convolution layers using 3x3 filters opening 32, 64, 128, and 256 with ReLU activation, followed by a poll layer with a 2x2 filter. The full connected layer uses 2 hidden layers with dropout layers. The training model was conducted with a 5-fold cross-validation method using epoch 100, 'adam' optimization algorithm with a learning rate of 0.0001, and batch size 10. The success of a model is seen based on the calculation of the confusion matrix. The test results showed that the accuracy rate of the third model using a combination of max polling and average polling performed best with an introduction accuracy of 98.75%, the first model used max polling with an accuracy of 98.25% while the lowest accuracy on the second model used average polling with an accuracy of 97.75%.