Discovering fully semantic representations via centroid- and orientation-aware feature learning.

Area of Science:

• Scientific image analysis
• Machine learning for scientific data

Background:

• Learning robust image representations is crucial for scientific discovery.
• Variations in object centroids and orientations pose challenges in scientific image analysis.

Purpose of the Study:

• To introduce a novel neural network, the centroid- and orientation-aware disentangling autoencoder (CODAE).
• To develop a method for learning meaningful, invariant image features across scientific domains.

Main Methods:

• Utilized an encoder-decoder neural network architecture.
• Incorporated a translation- and rotation-equivariant encoder with Euler encoding.
• Applied an image moment loss for feature extraction.

Main Results:

• CODAE successfully extracts features invariant to object positions and orientations.
• The model learns object centroids and orientations from randomly transformed images.
• High-quality aligned and exact view reconstructions of input images were achieved.

Conclusions:

• CODAE effectively addresses the challenge of learning invariant image representations in scientific domains.
• The method demonstrates versatility across life sciences, material science, and astronomy datasets.
• CODAE facilitates enhanced analysis of scientific images with positional and orientational variations.