Indonesian Traditional Cake Classification Using Convolutional Neural Networks

Abstract

In Indonesia, there are many types of cakes that are categorized as traditional snacks. Refer to the Kamus Besar Bahasa Indonesia; snacks are defined as foods that are peddled or mean bites. Snacks are classified based on how they are made, and some are based on the taste of the snacks. Traditional snacks are a part of Nusantara culture that is mandatory for those born and live in Indonesia to preserve them. But in reality, many people tend to consume and know more about modern snacks than traditional ones. In fact, not many people have even tried traditional snacks or even made their own at home. This application was developed to help people distinguish and recognize the various kinds of cakes on the market. With convolutional neural networks technology in machine learning, people can use image classification presented through mobile applications with accuracy above 90% for Indonesian traditional cake recognition.

The Conceptual Idea of Strategy to Increase Sales of Traditional Wet Cakes in Gunungkidul Regency

“…This is in line with the changing trend of tourists who are willing to come to an area to enjoy the local specialties, even willing to pay a high price for them (Rismiyanto & Danangdjojo, 2015). One type of traditional snack is "kue basah" or wet cakes (Karlita et al, 2022). Wet cakes is a type of cake that has a soft, moist, and tender texture (Anisa et al, 2022).…”

“…These cakes are usually made for traditional ceremonies, rituals, celebrations, and daily snacks (Pamungkas et al, 2023). Nowadays, most people are more familiar with modern snacks than traditional ones (Karlita et al, 2022). The main challenge for developing traditional wet cake into food souvenirs is their short shelf life (Anisa et al, 2022), and similar to other SMEs, marketing is a challenge (Hafsah, 2004).…”

Recognition of Indonesian Traditional Cakes using The MobileNet Algorithm

“…The output of the base model is fed into a pooling layer to get the average value and convert it to a mean or vector. Then the dropout on the fully connected layer is 0.5 to avoid overfitting [10]. There are two activation functions in the hidden layer, namely ReLU and Softmax, where a dropout of 0.5 is given between the two layers.…”

“…They are inspired by CNN breakthroughs and Deep learning algorithms in object detection, image classification, deep learning research and applications, and CNN in plant disease detection which is currently the center of research [9]. Similar research have been done before with various methods, such as from Tita Karlita, et al, who used MobileNetv2 method with 8 image labels to classify cake types and produced 90% accuracy [10], then there was also from Zaenal Abidin, et al using the method CNN -MobileNetv2 with 1545 images for 8 labels and produces an accuracy of 92.5% [11].…”