Neural Phase Correlation learns image alignment from scratch
New method outperforms on cardiac MRI without explicit encoding
A new paper from independent researcher Cole Reynolds proposes Neural Phase Correlation, a fundamental rethink of how computer vision models handle image correspondence. Traditional deep learning methods encode each image independently and let a decoder discover the mapping implicitly, treating the transformation as a byproduct rather than a first-class object. Phase correlation, the classic exception, measures the relationship directly in the Fourier domain but is limited to global translation due to its fixed basis. Reynolds' approach learns the basis itself, enabling the same algebraic machinery to handle dense non-rigid deformations and even unitary dynamics.
Results on medical imaging benchmarks are striking: on the ACDC cardiac-MRI dataset, Neural Phase Correlation matches or exceeds all prior published baselines in both registration directions. On CAMUS echocardiography, it achieves state-of-the-art accuracy without needing auxiliary scoring or adaptive-smoothness mechanisms that other methods rely on. Extending beyond vision, the framework also recovers the Hermite-function eigenstates and quantized energy levels of the 1D quantum harmonic oscillator—demonstrating that learning basis functions can uncover physical structure from observation pairs alone.
- Learns transformation basis directly from data, unlike classical phase correlation which is limited to global translation
- Matches or exceeds prior baselines on ACDC cardiac-MRI and achieves state-of-the-art on CAMUS echocardiography without auxiliary mechanisms
- Also recovers Hermite eigenstates and energy levels of quantum harmonic oscillator, showing cross-domain generality
Why It Matters
A principled alternative to deep registration that could reduce data needs and improve robustness in medical imaging and physics.