Comparison of CNN Transfer Learning in Detecting Superior Local Fruit Types in Bali

Abstract

Bali Province is an island that has unique geographical conditions, as well as the diversity of fruit it has. The specialty of local fruit is not only of economic value for food needs but also for religious ceremonial needs. Bali provincial government is currently actively promoting local fruit so that it can be used as consumption for Bali's increasingly rapid tourism. Several superior fruits were developed as an effort to raise the potential of local fruit in the tourism sector. Some of the superior fruits are Balinese snake fruit and sapodilla. However, snake fruit is one of the superior local fruits in Bali which has not experienced degradation over time. This research aims to detect the types of snake fruit in Indonesia. This fruit is not popular compared to imported fruit. Therefore, an application is needed to recognize this type of snake fruit automatically. This research uses a deep learning method with the CNN (Convolutional Neural Network) algorithm. This algorithm is able to recognize and classify an image well. The fruit images used were 400 fruits for 4 types of snake fruit. Where the training data for snake fruit is special because it has different skin and fruit contents. In this research, 2 transfer learning models from the CNN algorithm were also compared, namely mobilenetv2 and ResNet152. Based on the test results, it was found that the best level of accuracy was obtained using the ResNet152 model with an accuracy value of 92% in identifying images of Balinese snake fruit.