Recognition of Rare Microfossils Using Transfer Learning and Deep Residual Networks

Area of Science:

• Paleontology
• Artificial Intelligence
• Machine Learning

Background:

• Early Cambrian microfossils are key to understanding animal phyla origins.
• Rare specimens and low manual selection efficiency hinder micropaleontology research.

Purpose of the Study:

• To develop an efficient method for identifying specific taxa of Cambrian microfossils.
• To leverage artificial intelligence and pre-trained models for paleontological data analysis.

Main Methods:

• Utilized a 34-layer deep residual neural network, migrating an ImageNet pre-trained model.
• Froze low-layer parameters and retrained high-layer parameters for microfossil image recognition.
• Created training sets with varied numbers of randomly selected microfossil images per taxon.

Main Results:

• Achieved average recognition accuracy higher than 0.97 with 50 training images per taxon.
• Reached 0.85 accuracy with only three training samples per taxon.
• Outperformed prevalent methods like the transpose convolutional neural network (TCNN).

Conclusions:

• Demonstrated the feasibility of using natural image datasets (ImageNet) for training microfossil recognition models.
• Provided a promising AI tool for the discovery of rare Cambrian microfossils.
• Overcame limitations of traditional manual selection in micropaleontology.