Underwater imaging is essential for marine resource exploration. Compared with terrestrial environments, strong scattering and complex backgrounds in water impose stringent requirements on machine vision, including high sensitivity, broadband response, wide field of view, and efficient data preprocessing. Here, we demonstrate an oxide-semiconductor-based biomimetic fisheye vision system for high-accuracy underwater recognition. We design a sandwich-structured phototransistor array as an artificial retina, where high-mobility, high on/off ratio ITZO functions as the channel layers and narrow-bandgap SnO serves as the light-absorbing interlayer, forming an n-ITZO/p-SnO/n-ITZO architecture. Benefiting from this design, the phototransistor exhibits a broadband response from the ultraviolet to the near-infrared with an average detectivity >1011 cm Hz1/2 W-1. The n-p-n band configuration further enables effective control of photogenerated carrier separation and recombination, resulting in highly linear and stable pulse-driven weight modulation as well as pronounced short-term memory and temporal sensitivity. These characteristics allow the retinal sensor to operate as a physical reservoir, supporting neuromorphic and in-sensor computing. Meanwhile, we optimized the 3D curvature of the retina to compensate optical aberrations, enabling wide-field-of-view imaging with extremely low aberrations. Using this fully hardware-based framework with algorithms, the average recognition accuracy for underwater fish targets reached 90.86%, which remains at 80.15% under strong Gaussian noise and above 70% even under salt-and-pepper noise perturbations. This work establishes a compact and efficient hardware paradigm for next-generation underwater visual systems.
Zeng et al. (2026) studied this question.