The exploration of the transfer learning technique for Globotruncanita genus against the limited low-cost light microscope images


Ozer I., KARACA A. C., KAYA ÖZER C., Gorur K., Kocak I., Cetin O.

Signal, Image and Video Processing, 2024 (SCI-Expanded) identifier

  • Publication Type: Article / Article
  • Publication Date: 2024
  • Doi Number: 10.1007/s11760-024-03322-x
  • Journal Name: Signal, Image and Video Processing
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, zbMATH
  • Keywords: Deep learning, Globotruncanita Genus, Microfossil, Transfer learning
  • Yıldız Technical University Affiliated: Yes

Abstract

Microfossils are single-celled micro-organisms and are noted as a powerful analysis way in earth sciences for determining geological age and in paleoenvironmental studies. However, the accurate taxa of fossil species manually using a microscope requires considerable time and labor by domain experts with extensive knowledge and experience due to these organisms’ complex structure and morphology. Therefore, developing an automated system for this process is considered an important research area with low-resolution records. In this study, we have focused on Globotruncanita genus as species-level identification for three species using transfer learning-based pre-trained deep learning models over low-cost light microscope images. Each of the three species has reported complex morphological differences regarding paleoecological interpretations. Thus, it is a concerned very important task to differentiate the three species with a limited number of specimens. In this point, the transfer learning technique is crucial to employ the strengths of a pre-trained deep convolutional neural network (CNN) model with many learned filters by training on millions of images. As far as we know, this research study is the first attempt to explore the species-level identification of the Globotruncanita genus by implementing a transfer learning technique with many pre-trained deep models in the existing literature. The proposed microfossil prediction models have shown up to 96.66% accuracy and 0.978 AUC score.