Automated Quantification of Brittle Stars in Seabed Imagery Using Computer Vision Techniques

Area of Science:

• Marine Biology
• Benthic Ecology
• Image Analysis

Background:

• Underwater video surveys are crucial for marine benthic research, generating large datasets.
• Manual analysis of video transects and mosaic maps is time-consuming and labor-intensive.
• Automated image segmentation and quantitative evaluation are needed to address data volume and analysis time.

Purpose of the Study:

• To investigate the effectiveness of automated image segmentation and counting techniques for brittle stars in underwater mosaic maps.
• To assess the performance of a deep convolutional neural network combined with blob detection for brittle star quantification.
• To compare different annotation methods (disc vs. full shape) for efficiency in brittle star studies.

Main Methods:

• Utilized a deep convolutional neural network with pre-trained weights for image segmentation.
• Applied a blob detection technique for post-processing and counting segmented instances.
• Annotated mosaic maps containing hundreds of brittle star instances.
• Tested disc markers versus full shape masks for annotation efficiency.

Main Results:

• The segment-and-count approach demonstrated effectiveness in brittle star segmentation and counting.
• Disc annotations were found to be faster than full shape masks for brittle stars.
• Underwater image enhancement techniques did not significantly improve segmentation accuracy but may be useful for data augmentation.

Conclusions:

• Automated segmentation and counting using deep learning offer a promising solution for analyzing large underwater video datasets in marine benthic research.
• Disc markers are a recommended, efficient annotation strategy for brittle stars in this context.
• Further research into image enhancement for data augmentation could be beneficial.