MorphoCluster: Efficient Annotation of Plankton Images by Clustering

Area of Science:

• Marine Biology
• Data Science
• Computer Vision

Background:

• The increasing volume and complexity of marine data necessitate advanced interpretation tools.
• Current annotation methods struggle to keep pace with data growth, requiring more efficient solutions.

Purpose of the Study:

• To introduce MorphoCluster, a novel software tool for efficient and accurate annotation of large image datasets.
• To augment human capabilities in pattern discovery and object classification within large data volumes.

Main Methods:

• Embedding unsupervised clustering within an interactive annotation process.
• Aggregating similar images into data-driven clusters.
• Developing a tool that allows adaptive granularity in sorting schemes.

Main Results:

• MorphoCluster processed 1.2 million objects into 280 classes in 71 hours, achieving 16,000 objects per hour.
• Achieved high precision, with 90% of classes having a precision of 0.889 or higher.
• Demonstrated increased annotation consistency and throughput compared to human experts.

Conclusions:

• MorphoCluster offers a fast, accurate, and consistent solution for large-scale image annotation.
• The tool provides fine-grained, data-driven classification and enables novelty detection in datasets.
• It effectively addresses the growing challenges in interpreting complex marine imagery.