Obtaining high-quality images for training AI models in the field of plankton identification, particularly cyanobacteria, is a challenging and time-critical task that necessitates the expertise of biologists. Data augmentation techniques, including conventional methods and GANs, can improve model performance, but GANs typically require large training datasets to produce high-quality results. To tackle this issue, we employed the StyleGAN2ADA model on a dataset of 9 cyanobacteria genera plus non-cyanobacterial microalgae. We evaluated the generated images using both qualitative and quantitative metrics. Qualitative assessments involved a psychophysical test conducted by three expert biologists to identify shape and texture deviations or chromatic aberration that might impede visual classification. Additionally, three non-reference image quality metrics based on perceptual features were used for quantitative assessment. Images meeting quality standards will be incorporated into classification models to assess the performance improvement compared to the original dataset. This comprehensive evaluation process ensured the suitability of generated images for enhancing model performance.

This work was funded by project TED2021-132147B-100 (funded by MCIN/AEI/10.13039/501100011033 and by the European Union NextGenertionEU/PRTR).

17 pags., 11 figs., 5 tabs. -- Event: SPIE Photonics Europe, 2024, Strasbourg, France

Peer reviewed