wenmin-wu/timeseries-correlated-double-sampling

Correlated Double Sampling (CDS)

Overview

For sensors with paired read/reset cycles, subtract even frames from odd frames (or vice versa) to cancel common-mode noise (kTC noise, bias drift, readout offsets). Halves the frame count but dramatically improves SNR. Common in astronomical detectors, IR sensors, and scientific imaging.

Quick Start

import numpy as np

def correlated_double_sampling(signal):
    """Subtract even (reset) frames from odd (read) frames.
    
    Args:
        signal: (..., T, ...) with T frames in read/reset pairs
    Returns:
        (..., T//2, ...) CDS-processed signal
    """
    return signal[..., 1::2, :] - signal[..., ::2, :]

# Often followed by temporal binning
def cds_and_bin(signal, bin_size=10):
    cds = correlated_double_sampling(signal)
    n_bins = cds.shape[-2] // bin_size
    binned = np.zeros((*cds.shape[:-2], n_bins, *cds.shape[-1:]))
    for i in range(n_bins):
        binned[..., i, :] = cds[..., i*bin_size:(i+1)*bin_size, :].mean(axis=-2)
    return binned

Key Decisions

• Odd - even: convention depends on sensor — check if read or reset comes first
• Pair binning after CDS: sum/mean consecutive CDS frames to further reduce noise by √N
• Spatial mean before CDS: for 2D detectors, average spatial pixels first if only temporal signal matters

References

• Source: ariel-only-correlation
• Competition: NeurIPS - Ariel Data Challenge 2024