Learning Transformation-Invariant Local Descriptors With Low-Coupling Binary Codes.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Binary local descriptors are crucial for image analysis but struggle with geometric transformations and correlated bits.
• Existing methods lack robustness to transformations and effective strategies for handling highly correlated bits from image hashing.

Purpose of the Study:

• To propose an unsupervised learning method for transformation-invariant binary local descriptors (TBLD).
• To address limitations of existing descriptors, specifically vulnerability to geometric changes and bit correlation issues.

Main Methods:

• TBLD employs simultaneous projection into high-dimensional and low-dimensional spaces for transformation invariance.
• An Adversarial Constraint Module with Wasserstein loss is used to reduce bit correlations by guiding descriptor learning with low-coupling binary codes.
• Enforces distinct binary local descriptors for dissimilar image patches.

Main Results:

• Experimental results on three benchmark datasets demonstrate the superiority of TBLD.
• The proposed method outperforms existing state-of-the-art techniques in descriptor learning.
• Achieved enhanced robustness against geometric transformations and reduced bit correlations.

Conclusions:

• The developed TBLD method effectively addresses key limitations in binary local descriptor learning.
• TBLD offers a robust and efficient approach for image analysis tasks requiring invariant descriptors.
• The unsupervised learning strategy provides a promising direction for future research in binary descriptor design.