Beyond photon detection: photodetectors as cognitive pixels in machine vision

Machine vision is increasingly limited by the physical separation between photon capture, memory, and computation, a bottleneck visible in conventional CMOS imagers whose energy and latency are dominated by data movement rather than visual inference. Against this backdrop, this review reframes recent progress in optoelectronic memristors, photosynapses, in-sensor processors, spectral pixels, tunable photodetectors, and photonic physically unclonable devices as a broader transition in which the photodetector evolves from a passive photon to electron transducer into an information processing element. We organize the literature around three coupled questions, namely what these devices can now perform, what fundamental limitations still prevent deployment beyond silicon based digital vision, and what design strategies may overcome them. Across platforms ranging from van der Waals semiconductors, halide perovskites, and wide-bandgap oxides to III nitride heterostructures, ferroelectric and phase change materials, organic systems, and bioinspired composites, a common trajectory emerges. Progress is no longer defined by isolated synaptic curves and record responsivities but by four coupled requirements, that is, a measurable internal state variable, a verified in-device visual primitive, a quantified system-level advantage, and a credible path to scalable back-end-of-line integration, together defining a CMOS integrated cognitive pixel for next-generation machine vision.