Consistency-Driven Dual-Teacher Framework for Semi-Supervised Zooplankton Microscopic Image Segmentation.

Area of Science:

• Marine biology
• Computer vision
• Ecological research

Background:

• Accurate marine biodiversity assessment relies on understanding marine ecosystems.
• Semantic segmentation of marine species is vital for ecological studies.
• Microscopic zooplankton image segmentation is challenging due to complex morphology and limited annotated data.

Purpose of the Study:

• To develop an effective semi-supervised learning framework for microscopic zooplankton segmentation.
• To address limitations of existing methods in handling variable morphologies and scarce annotations.
• To improve the accuracy and reliability of zooplankton image analysis for marine biodiversity research.

Main Methods:

• A consistency-driven dual-teacher framework utilizing two heterogeneous networks (global and local feature focus).
• Dynamic fusion-based pseudo-label filtering strategy integrating hard and soft labels based on consistency and confidence.
• Implementation and evaluation on the self-constructed Zooplankton-21 Microscopic Segmentation Dataset (ZMS-21).

Main Results:

• The proposed dual-teacher framework significantly outperforms existing semi-supervised segmentation methods.
• Achieved mean Intersection over Union (mIoU) scores of 64.80% (1/16), 69.58% (1/8), 70.32% (1/4), and 73.92% (1/2) labeled data.
• Demonstrated consistent performance improvement across various annotation ratios.

Conclusions:

• The consistency-driven dual-teacher framework offers a robust solution for challenging microscopic zooplankton segmentation.
• The method effectively leverages limited annotations through complementary supervision and adaptive pseudo-labeling.
• This approach advances automated analysis of marine plankton, crucial for marine biodiversity monitoring and conservation efforts.