A Metric-Based Few-Shot Learning Method for Fish Species Identification with Limited Samples.

Area of Science:

• Marine Biology
• Computer Science
• Artificial Intelligence

Background:

• Accurate fish species identification is crucial for marine fisheries management and biodiversity monitoring.
• Limited datasets and visual similarities between species pose significant challenges in open-water environments.
• Existing methods struggle with identifying fish species effectively, especially when sample sizes are small.

Purpose of the Study:

• To develop an effective few-shot learning approach for fish species identification.
• To address the scarcity of marine fish datasets and the challenge of visually similar species.
• To improve the accuracy and efficiency of fish classification in complex scenarios.

Main Methods:

• A novel few-shot learning model incorporating an embedding module and a metric function.
• The embedding module leverages species distribution in embedding space to differentiate similar phenotypes.
• The metric function enhances classification performance and addresses limited sample sizes.

Main Results:

• The proposed model achieved a 2% to 10% improvement in accuracy compared to prototypical networks.
• Effective identification of fish species was demonstrated even with small sample sizes and in complex environments.
• The model was trained end-to-end on public datasets: Croatian fish, Fish4Knowledge, and WildFish.

Conclusions:

• The developed few-shot learning approach offers a valuable technological tool for marine fisheries resource exploration.
• This method significantly contributes to the preservation of fish biodiversity by enabling accurate species identification.
• The approach effectively overcomes limitations of small sample quantities and complex visual scenarios in fish identification.