Overfitting Reduction Approach in MobileNet for Junk Image Classification

Abstract

The challenge faced in waste management, especially in waste image classification, is the low accuracy of the model which can cause errors in sorting waste types, thus reducing the effectiveness of the recycling process. This study aims to develop a more accurate waste image classification model using the MobileNet approach, which is enhanced with overfitting reduction techniques such as data augmentation, dropout, and L2 regularization. The methodology used is experimental research. The first model was built without overfitting reduction techniques, while the second model applied augmentation, dropout, and L2 regularization techniques. The training results of the first model showed a high training accuracy of up to 99.95%, but there was a decrease in validation accuracy of up to 80%, indicating overfitting. In contrast, the second model managed to maintain a more stable and higher validation accuracy, with a validation accuracy reaching 91.24% at the end of the epoch, indicating an increase in the model's generalization ability to unseen data. These findings contribute to improving the accuracy and efficiency of waste sorting, which supports better waste management and recycling systems, as well as the development of machine learning models in waste and environmental management.