Analisis Klasifikasi Citra Daging Sapi dan Daging Babi Menggunakan VGG16, EfficientNetB0, dan InceptionV3

Abstract

Konsumsi daging di Indonesia semakin meningkat seiring dengan terus meningkatnya jumlah penduduk, pendapatan, dan konsumsi protein hewani. Pengidentifikasian sumber daging menjadi isu yang sensitif dan penting terkait dengan keamanan dan jaminan produk makanan sehingga tindakan pemalsuan daging dapat dihindari. Penelitian ini menggunakan model Convolutional Neural Network (CNN) dengan arsitektur EfficientNetB0, VGG16, dan InceptionV3. Data yang digunakan berupa citra daging sapi dan daging babi yang berjumlah 500 citra. Beberapa hyperparameter diuji untuk mendapatkan performa model yang paling optimal seperti freeze rate 50% dan 75%, jumlah dense layer yang digunakan, dan learning rate 0,001 dan 0,0001. Hasil penelitian menunjukkan arsitektur EfficientNetB0 belum mampu membedakan citra daging sapi dan daging babi dengan validation accuracy sebesar 64,58% dan testing accuracy sebesar 50%. InceptionV3 dengan freeze rate 75%, learning rate 0,0001, dan 1 dense layer merupakan model paling optimal dengab validation accuracy sebesar 95,83% dan testing accuracy sebesar 75%. Sementara itu, VGG16 memiliki validation accuracy yang lebih rendah sebesar 93,75% dan testing accuracy sebesar 65%.

Meat consumption in Indonesia is continuously increasing in line with the increase of population, income, and animal protein consumption. Identifying the source of meat has become a sensitive and important issue related to the safety and assurance of food products in order to avoid meat adulteration. This research uses the Convolutional Neural Network (CNN) model with EfficientNetB0, VGG16, and InceptionV3 architecture. The data used are images of beef and pork, with 250 images each. Several hyperparameters were tested to obtain the most optimal model performance such as 50% and 75% freeze rate, number of dense layers used, and learning rate 0,001 and 0,0001. The results indicate that EfficientNetB0 architecture has not been able to distinguish between beef and pork images with a validation accuracy of 64,58% and testing accuracy of 50%. InceptionV3 with a freeze rate of 75%, learning rate of 0,0001, and 1 dense layer becomes the most optimal model with a validation accuracy of 95,83% and testing accuracy of 75%. Meanwhile, VGG16 has a lower validation accuracy of 93,75% and testing accuracy of 65%.