Effect of Pre-processing Dataset on Classification Performance of Deep Learning Model for Detection of Oil Palm Fruit Ripe

Abstract

Palm oil is Indonesia's main commodity with the highest production in the world, merely in the process of determining the maturity of palm oil, it is still processed manually so that the results of the quality of palm oil are not optimal. This paper presents the deep learning model analysis to determine the maturity level of oil palm fruit accurately. In this study, data collection of oil palm fruit images was carried out in oil palm plantations, with six categories of maturity levels, namely ripe, underripe, overripe, immature, empty and abnormal. The processing of the dataset begins with the process of removing the background to focus on the processed palm oil image, then augmentation with various angles of rotation is carried out to increase the dataset. The development of a classification model for oil palm fruit maturity using deep learning, namely: Alexnet, MobileNetV1 and MobileNetV2. Furthermore, an evaluation is carried out using the confusion metric to compare the performance of the best classification model. The result of model testing shows MobileNetV1 has the highest performance among the three tested models. Thus, it can be concluded that the pre-processing of the dataset could improve the performance of the MobileNetV1 model compared to MobileNetV2.