Research map / Detection and segmentation / Open-set and few-shot detection
Selective Nonlinearities Removal from Digital Signals
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains two U-Net-inspired networks that take 2D amplitude stacks of Fourier transforms (under quadratic/cubic phase-correction factors) and output nonlinearity-free Fourier amplitudes, removing either second-order or third-order nonlinearity from OCT-style signals selectively.
Problem
Fourier-transform-based signals (OCT, MRI, FTS) are distorted by second- and third-order nonlinearities in the modulation, broadening or distorting peaks and degrading image resolution; existing linearisation removes all orders together and cannot isolate one order.
Approach
Build a 32x1024 stack of Fourier amplitudes of the signal multiplied by different quadratic (or cubic) phase factors, then train a U-Net-style network (with max-pooling to 1D output) to stitch the nonlinearity-free regions into a Fourier amplitude free of only that nonlinearity order.
Key results
- Both networks reach mean GoF of 99.98% at a 1% error threshold on the synthetic test set.
- At a 0.1% error threshold, 9.52% (Network #1) and 3.07% (Network #2) of test predictions fall below 95% GoF.
- On experimental OCT data, Network #1 sharpens peaks comparably to traditional linearisation, with a higher peak in the mirror test; inference takes 39 ms per stack on a Quadro RTX 6000 GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove second-order nonlinearity while preserving third-order nonlinearity | Network #1 · Synthetic test dataset | GoF (1% error threshold): 99.98% | Helped |
| Remove third-order nonlinearity while preserving second-order nonlinearity | Network #2 · Synthetic test dataset | GoF (1% error threshold): 99.98% | Helped |
| Use Network #1 to remove second-order nonlinearity | OCT mirror signals at different depths · Mirror at 11 different depths | Peak width | Helped |
| Use Network #2 to remove third-order nonlinearity | OCT mirror signals at different depths · Mirror at 11 different depths | Peak elongation | Helped |
| Use Network #1 instead of traditional linearisation | OCT mirror signals · Mirror at different depths | Peak shape | Helped |
| Use Network #1 to remove second-order nonlinearity | OCT signals · Sapphire glass | Image quality | Helped |
| Use Network #2 to remove third-order nonlinearity | OCT signals · Sapphire glass | Image quality | No clear effect |
| Use Network #1 instead of traditional linearisation | OCT signals · Grape and cucumber | Image quality | Mixed |
| Use traditional linearisation instead of Network #1 | OCT signals · Mirror at different depths | Peak shape | Hurt |
Surprising findings
- Network #1 output retains a third-order 'tail' at depth, which traditional linearisation removes, suggesting the traditional method is not perfect.
- Network #2 gave no visually significant improvement on the glass and biological data because third-order nonlinearity there is very small.
Limitations
- Predictions produce small non-zero values where ground truth is exactly zero, lowering GoF.
- Trained predominantly on high signal-to-noise signals, so low-SNR cucumber data loses detail.
- Processing speed is insufficient for real-time OCT imaging.
Experimental setup
Datasets: computer-generated Gaussian-shaped OCT-like signals with 2-12 interfaces experimental OCT data of mirror, sapphire glass, grape and cucumber
Benchmarks: held-out synthetic test set of 14,993 objects experimental OCT mirror, glass, grape, cucumber data
Metrics: Goodness-of-Fit Approximity (GoF) mean absolute error
Baselines: traditional OCT linearisation method [1] machine learning nonlinearity removal [9]
Builds on: U-Net Optuna
Closest papers by meaning
- Step-Calibrated Diffusion for Biomedical Optical Image Restoration (AAAI 2025 · similarity 0.80)
- All-Optical Nonlinear Diffractive Deep Network for Ultrafast Image Denoising (CVPR 2025 · similarity 0.79)
- CosAE: Learnable Fourier Series for Image Restoration (NeurIPS 2024 · similarity 0.79)
- Quality-Improved and Property-Preserved Polarimetric Imaging via Complementarily Fusing (NeurIPS 2024 · similarity 0.78)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.78)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.