Fig. 1. (Color online) (a) Schematic of traditional frame-by-frame detecting system. Detector genetrates output for subsequent computing in every single frame. (b) Schematic of multiframe-integrated, in-sensor computing using persistent photoconductivity effect. Detector continuously detects multiple frames and only generates one final output state for analysis, which already memorizes the information of past and current frames, spatially and temporally. The final state is input to the subsequent linear classifier, serving as the readout layer.

Fig. 2. (Color online) (a) Schematic image of retinomorphic MoS₂ photodetector. (b) Scanning electron microscopy (SEM) imaging (left) and wavelength-dispersive X-ray spectroscopy (WDS) imaging (sulfur element and molybdenum element image are present in red and green, respectively) of retinomorphic MoS₂ photodetector, Scarbar, 50 μm. (c) Raman spectra of retinomorphic MoS₂ photodetector. The inset shows the transmission electron microscopy (TEM) images of cross-sectional MoS₂ flake, Scarbar, 10 nm. (d) X-ray photoelectron spectroscopy (XPS) of the retinomorphic MoS₂ photodetector. The inset shows the optical image of a 1 cm × 1 cm retinomorphic MoS₂ photodetector array.

Fig. 3. (Color online) (a) I−V characterization of retinomorphic MoS₂ photodetector in logarithmic scale. The inset shows the I−V curve on a linear scale. (b) The persistent photoconductivity effects observed in retinomorphic MoS₂ photodetector illuminated under laser pulses (520 nm, 10 mW). Pink rectangle: light on; blue rectangle: light off. (Insert: Photocurrent of Au/MoS₂/Au device were measured under illumination by light pulses with different power (λ = 520 nm, 3, 5, 10, 12, 14 mW laser power)). (c) 3-bit light pulse inputs ranging from "000" to "111", each with a pulse width and interval of 100 and 900 ms were used. (d) The resultant normalized photocurrent characteristics, including input−output feature extraction, were analyzed using a retinomorphic MoS₂ photodetector.

Fig. 4. (Color online) (a) Schematic of the mission proposed, target can move in two directions (clockwise/anticlockwise), we use light pulse irradiated onto detector array (8 × 8 pixels in our simulation) to represent the location of target. (b) Schematic of four heatmaps of photocurrent of all detectors after every single frame, the green one refers to clockwise, the blue one refers to anticlockwise, darker a pixel is, later a light pulse appears here. (c) Evolution of the accuracy rates based on multiframe-integrated RC system and traditional FC network within 100 epochs.