Fig. 4: Fingerprint recognition based on hardware DUV in-sensor RC system.

a Schematic of the proposed fully-hardware photoelectronic RC system for in-sensor fingerprint recognition, including photo-synapse reservoir layer which generates feature outputs, and memristor readout layer which performs network training. b Preprocessing method of the fingerprint images, including cropping, compressing, binarizing, and rejoining. c The evolution of the accuracy rates based on single and dual features during readout network training. The training process with dual features demonstrates a much faster convergence. d The colormaps and e statistic histograms of the 40 × 5 weights of the simulation and hardware experiment, respectively. The actual conductance values read from hardware were multiplied by a constant of 1.25 × 10⁴ for better comparison with the simulated weights. f Influence of stochastic noise on recognition accuracy rates for fingerprint recognition of the RC system, implemented by full-precision simulation, limited-precision simulation, and hardware experiment, respectively.