Optimization Strategy on Deep Learning Model to Improve Fruit Freshness Recognition

Abstract

The high fruit production during the harvest season is a challenge in the process of sorting fresh fruit and rotten fruit in plantations. Automatic fruit freshness classification based on deep learning can speed up the sorting process. However, building a model with high accuracy requires the right strategy based on the dataset's characteristics. This research aims to apply optimization strategies to deep learning models to improve model performance. The optimization strategy is implemented by optimizing the model using fine-tuning strategy by selecting the best parameters based on learning rate, optimizers, transfer learning, and data augmentation. The transfer learning process is applied based on the dataset's characteristics by training some parameters with a size of 30% and 60%, which were tested in four scenarios. The fine-tuning strategy is applied to three Deep Learning models, i.e., MobileNetv2, ResNet50, and InceptionResNetV2, which have various parameter sizes. Based on test results, fine-tuning strategy produces the best performance up to 100% with a learning rate of 0.01, the SGD optimizers on the InceptionResNetV2 model are trained on 60% of the parameters.