Fine-Grained Identification of Benthic Diatom Scanning Electron Microscopy Images Using a Deep Learning Framework
Benthic diatoms are key primary producers in aquatic ecosystems and sensitive bioindicators for water quality monitoring; for example, the Yellow River Basin exhibits high diatom species diversity. However, traditional microscopic identification of such species remains inefficient and inaccurate. To...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-05-01
|
| Series: | Journal of Marine Science and Engineering |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2077-1312/13/6/1095 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Benthic diatoms are key primary producers in aquatic ecosystems and sensitive bioindicators for water quality monitoring; for example, the Yellow River Basin exhibits high diatom species diversity. However, traditional microscopic identification of such species remains inefficient and inaccurate. To enable automated identification, we established a benthic diatom dataset containing 3157 SEM images of 32 genera/species from the Yellow River Basin and developed a novel identification method. Specifically, the knowledge extraction module distinguishes foreground features from background noise by guiding spatial attention to focus on mutually exclusive regions within the image. This mechanism allows the network to focus more on foreground features that are useful for the classification task while significantly reducing the interference of background noise. Furthermore, a dual knowledge guidance module is designed to enhance the discriminative representation of fine-grained diatom images. This module strengthens multi-region foreground features through grouped channel attention, supplemented with contextual information through convolution-refined background features assigned low weights. Finally, the proposed method integrates multi-granularity learning, knowledge distillation, and multi-scale training strategies, further improving the classification accuracy. The experimental results demonstrate that the proposed network outperforms comparative methods on both the self-built diatom dataset and a public diatom dataset. Ablation studies and visualization further validate the efficacy of each module. |
|---|---|
| ISSN: | 2077-1312 |