Comparison of Transfer Learning Architecture Performance for Indonesian Auction Object Classification

Comparison of Transfer Learning Architecture Performance for Indonesian Auction Object Classification

image classification;indonesian auction;PNBP; transfer learning

The Indonesian auction, one of the sources of Indonesia's income for Non-Tax State Revenue (PNBP), faces challenges in accurately classifying auction objects, limiting revenue optimisation. This research aims to compare the performance of several transfer learning architectures on the Indonesian Auction Object Dataset, which includes categories such as Buildings, Cars, Motorbikes, and Salvage Materials. Seven pre-trained transfer learning models—MobileNetV2, NASNetMobile, EfficientNetV2B0, DenseNet121, Xception, InceptionV3, and ResNet50V2—were evaluated against a baseline model, focusing on validation accuracy, model size, and computational efficiency. MobileNetV2, NASNetMobile, DenseNet121, Xception, InceptionV3, and ResNet50V2 all achieved 100% validationaccuracy, outperforming the baseline model's 96.5% accuracy. MobileNetV2 stands out for its efficiency, reaching 100% accuracy in just eight epochs with a compact model size of 11.1 MB. In contrast, EfficientNetV2B0 performed poorly on this dataset, achieving only 25% validation accuracy. These findings confirm that transfer learning architectures can significantly improve auction object classification accuracy while reducing the model size and training time, highlighting the benefit of transfer learning for optimising Indonesian auction systems

Hanif Noer Rofiq1*

https://jurnal.iaii.or.id/index.php/RESTI/article/view/6082/1003

Directorate Transformation and Information Systems, Directorate General of State Assets Management, Ministry of Finance,

19-01-2025

FAJAR BAGUS W

PDF

ENGLISH

TEXT