The Impact of Data Augmentations on Deep Learning-Based Marine Object Classification in Benthic Image Transects.

Area of Science:

• Marine Biology
• Computer Vision
• Machine Learning

Background:

• Computer vision and data augmentation are crucial for analyzing large image datasets.
• Traditional methods struggle with marine science data due to low volume, class imbalance, and high annotation costs.
• Increasingly large image collections from remote marine habitats necessitate automated analysis for biodiversity assessment.

Purpose of the Study:

• To investigate the effectiveness of data augmentation for taxonomic classification in underwater benthic images.
• To compare standard data augmentation techniques with novel approaches tailored for marine imagery.
• To identify optimal data augmentation strategies for small, imbalanced marine datasets.

Main Methods:

• Evaluation of established geometric and photometric data augmentation techniques on marine image collections.
• Development and proposal of new data augmentation combination policies.
• Comparison of proposed policies against the AutoAugment algorithm for marine taxonomic classification.

Main Results:

• Standard data augmentation methods show varied performance on marine images, with some negatively impacting learning.
• Proposed data augmentation policies demonstrate superior performance compared to AutoAugment on marine image datasets.
• Effectiveness of data augmentation is highly dependent on the specific characteristics of marine imagery.

Conclusions:

• Data augmentation for marine computer vision requires domain-specific strategies.
• Tailored augmentation policies, incorporating background knowledge, are essential for small marine datasets.
• Optimized data augmentation significantly enhances taxonomic classification accuracy in underwater environments.