Memristor-based analog noise correction for infrared sensors

Feb 1, 2026 by X. Huang 黄, Peiwen 霈文 Tong 童, Q. Li 李, Tuo 拓 Ma 马, S. Han 韩, W. Wang 王, Y. Sun 孙 (Chinese Physics B)

DOI 10.1088/1674-1056/adecf9

We implemented a compact UNet inside a 1T1R memristor array to do analog-domain denoising and defective-pixel correction for IR sensors, training the network with measured device variability so the chip actually tolerates real memristor quirks. The result is a low-power, low-latency front-end that boosts recognition accuracy by ~50% versus uncorrected inputs while slashing ADC/processing overhead — a practical step toward in-sensor neural correction.

source S2, crossref