In-Sensor Visible and Mid-Infrared Dual-Band Perception and Computing

Machine vision in complex environments requires hardware that can integrate multispectral sensing with real-time front-end information processing. Visible light (VIS) and mid-infrared (MIR) signals provide complementary scene information, yet existing photodetectors cannot directly support VIS-MIR dual-band image processing at the sensor front end because their outputs are difficult to make both spectrally distinguishable and computation-compatible. Here, we report an electrically tunable lateral back-to-back dual-heterojunction device based on multilayer graphene/tungsten diselenide/black phosphorus for VIS-MIR dual-band in-sensor computing. The device exhibits electrically defined photoresponse states with spectrally distinguishable and bipolar characteristics, enabling simultaneous pattern recognition and spectral-band identification directly at the sensor level with 100% accuracy across three letter classes. It further supports dual-band in-sensor convolution in MIR-suppression and MIR-enhancement modes, allowing task-oriented modulation of MIR information while extracting VIS features. In addition, the device enables effective sensor-level suppression of VIS glare superimposed on MIR targets, improving downstream recognition accuracy from 33 to 98%. This work provides a device-level route toward task-oriented VIS-MIR information processing and efficient multispectral machine vision in complex environments.