In the noise spectrum of IR detectors, if thermal and photon noise are the high-pitched "white noise", 1/f noise is a low, stubborn, hard-to-dispel hum. It doesn't set the absolute sensitivity limit, yet directly governs the long-term stability, uniformity and low-frequency reliability of images — a stubborn enemy that high-performance cameras, especially those requiring quantitative thermometry and long monitoring, must conquer.
01 Identification: the "pink noise" that fights frequency
1/f noise, also called flicker or pink noise, shows its signature in the power spectral density: S(f) ∝ 1/f^γ, γ typically 0.8~1.3, ideally ~1. This means: noise power rises inversely as frequency drops; at very low frequencies 1/f can far exceed flat white noise (thermal, photon) and dominate. Engineering manifestations: temporal "drift" — a single pixel's output wanders slowly and aperiodically; spatial "fixed-pattern noise" — because pixels differ slightly in 1/f behavior, low-frequency or DC signals show a fixed spatial pattern that badly hurts uniformity; integration-time dependence — when long integration raises SNR, low-frequency 1/f mixes in more and can cancel the gain.
02 Physical origin: a micro-defect symphony of materials and interfaces
Unlike the "intrinsic" thermal noise, 1/f is a "excess" noise rooted in disorder, impurities, defects and imperfect interfaces. Mainstream models: carrier-number fluctuation — carrier count fluctuates via random trapping/release at defect levels (widely accepted for semiconductors); mobility fluctuation — carrier mobility fluctuates randomly from scattering, causing conductance fluctuation. In IR detectors specifically: bulk defects (grain boundaries, dislocations, impurities in VO?, a-Si films); metal-semiconductor contact noise (interface state-density fluctuation at electrodes — a hotspot); surface-state noise (dangling bonds and defect states on the sensitive surface contribute significantly). Key formula for a resistive material: V_f2 = (α_H V_b2 Δf) / (N_c f), where α_H is the Hooge parameter, V_b the bias voltage, N_c the total carrier count, f the frequency. Lowering 1/f means lowering α_H or raising N_c.
03 Concrete harm to IR imaging systems
The damage is systematic. First, it breaks non-uniformity correction: factory NUC builds gain/offset per pixel, but 1/f makes response (especially offset) drift slowly, so coefficients "expire" and fixed-pattern noise (ghosts, stripes) returns, forcing frequent recalibration. Second, it hurts long-integration and weak-signal detection: in astronomy or early-warning with extreme integration times, low-frequency 1/f accumulates and becomes the dominant SNR limiter that simple longer integration can't fix. Third, it lowers thermometry accuracy and stability: for high-precision industrial or scientific measurement, low-frequency pixel drift is misread as temperature change, adding system errors that ordinary filtering can't remove.
04 Suppression: a total encirclement from process to algorithm
Material & process (root cure): lower α_H via better growth (sputtering, CVD) for higher crystallinity and lower defect density; precisely control VO? oxygen content and crystallization; improve ohmic contact and interface passivation to cut surface-state noise; within limits enlarge the pixel sensing volume to raise N_c and proportionally lower 1/f — but this conflicts with the small-pixel, high-resolution trend.
ROIC design (circuit tricks): Correlated Double Sampling (CDS) — sample at integration start and end, subtract — perfectly cancels the constant low-frequency noise and offset, keeping only signal change; chopper stabilization shifts the low-frequency signal to high frequency then demodulates, avoiding the 1/f band; differential/bridge designs pair a sensing pixel with an optically shielded blind pixel so common-mode 1/f and drift are largely rejected in the differential amplifier.
Algorithm (software compensation): adaptive NUC updates coefficients dynamically from scene statistics or calibration references to track and compensate 1/f drift (constant-statistics, neural-network methods); temporal high-pass filtering on each pixel's time series removes slow drift but also discards real low-frequency temperature changes — use with care.
1/f noise, this low-frequency phantom rooted in the micro world of materials, is the last stubborn barrier between ideal uniform images and reality. Fighting it is a silent, enduring war from atomic-scale process to system-level algorithm. A trustworthy thermal image needs not only high peak SNR but stability and uniformity that withstand time. Conquering 1/f is the key step toward that reliability — making every degree of temperature measurement a faithful record of time, not just an instant capture.