Progress in in-sensor computing and applications based on photodetectors of two-dimensional materials

This paper provides a comprehensive review of recent advances in high-performance photodetectors based on two-dimensional materials and in-sensor computing for intelligent image processing, aiming to address the challenges of the “memory wall” and “power wall” caused by the separation of sensing, storage, and computing in traditional image sensors. Traditional image processing relies on the von Neumann architecture, where a large volume of raw data generated at the sensing end must be transmitted to independent computing units or cloud platforms for processing, leading to high energy consumption, significant latency, bandwidth burden, and security risks. Owing to their atomic thickness, high carrier mobility, weak short-channel effects, and tunable optoelectronic properties, two-dimensional (2D) materials provide an ideal physical foundation for achieving function integration of perception and computation. This paper discusses the topic from three perspectives: optical signal perception, image preprocessing, and advanced image processing. In terms of optical signal perception, 2D materials and their heterostructures exhibit ultrahigh responsivity, broadband operation, and fast response in light-intensity detection, enable miniaturized spectrometers through bandgap modulation and computational spectroscopy, and achieve compact, full-polarization analysis via twisted layers and metasurface structures. In image preprocessing, 2D material devices can perform convolution and feature extraction at the sensing end through linear photoresponse, suppress noise and extend dynamic range via superlinear and sublinear responses, and mimic biological visual adaptation in spectral and polarization domains to enhance image quality and robustness. In advanced image processing, the tunable photoresponse and memristive characteristics of 2D materials enable sensor-level integration of sensing, storage, and computation, This allows for the realization of matrix-vector multiplication and convolution operations within convolutional neural networks, significantly reducing power consumption and improving efficiency. Meanwhile, by implementing spike-rate and temporal encoding of optical signals in spiking neural networks, 2D material devices can achieve event-driven image recognition and classification under low-power and low-latency conditions. Furthermore, this paper highlights the challenges faced by 2D material image sensors, including scalable fabrication, heterogeneous integration with silicon technology, array- and circuit-level optimization, environmental stability and encapsulation, and system-level implementation, while envisioning their broad application prospects in intelligent imaging, wearable electronics, autonomous driving, and biomedical diagnostics. It is concluded that with the joint progress in materials science, device engineering, and artificial intelligence, 2D materials are expected to drive the development of next-generation low-power, high-performance, intelligent image processing platforms, and to become an essential foundation for future information perception and processing technologies.