Research Article
Research on Fine-Grained Classification of Small Sample Marine Organism Images
@INPROCEEDINGS{10.1007/978-3-031-65123-6_17, author={Huibin Luo and Zixin Lin}, title={Research on Fine-Grained Classification of Small Sample Marine Organism Images}, proceedings={Quality, Reliability, Security and Robustness in Heterogeneous Systems. 19th EAI International Conference, QShine 2023, Shenzhen, China, October 8 -- 9, 2023, Proceedings, Part II}, proceedings_a={QSHINE PART 2}, year={2024}, month={8}, keywords={Marine organisms Fine-grained classification Transfer learning Multi-model weighted ensemble Image augmentation}, doi={10.1007/978-3-031-65123-6_17} }- Huibin Luo
Zixin Lin
Year: 2024
Research on Fine-Grained Classification of Small Sample Marine Organism Images
QSHINE PART 2
Springer
DOI: 10.1007/978-3-031-65123-6_17
Abstract
Marine organisms exhibit high species diversity and minimal intra-species differences, often requiring the observation of local variations to accurately identify their categories. However, due to the complexity of underwater imaging conditions, marine organism image datasets are scarce, making accurate classification challenging. To address the problem of limited marine organism image datasets and the difficulty of fine-grained classification, we propose an image classification method based on transfer learning and multi-model ensemble. Firstly, a series of image augmentations are applied to the small sample marine organism images. Next, transfer learning is employed by leveraging the ImageNet-1000 pre-trained weights to aid rapid model learning. ResneSt and Efficientnet deep learning networks are selected for model training. Subsequently, a multi-model weighted ensemble method is used for marine organism image classification. Finally, model predictions are further optimized through binary classification. Experimental results demonstrate that the accuracies of the four models, ResneSt269, Efficientnet-b5, Efficientnet-b6, and “Efficientnet-b5 + Efficientnet-b6,” reach 97.5%, 97.68%, 97.86%, and 98.04%, respectively. After model optimization, the ensemble model “Efficientnet-b5 + Efficientnet-b6” achieves an accuracy of 99.11%. Therefore, the proposed method based on transfer learning and multi-model ensemble shows promise for fine-grained classification of small sample marine organism images.
