Color and orientation feature-attention sensors for efficient recognition

A key challenge in visual search is that target objects defined by specific features are often mixed with distractors. This demands intelligent vision systems capable of performing selective feature acquisition instead of randomly or exhaustively scanning the environment. Here, we report a feature-attention vision sensor that integrates a semantic and geometric “dual-attention” mechanism directly at the hardware level. The sensor features reconfigurable spectral responsivity from the visible to near-infrared range, enabling tunable color attention with a spectral rejection ratio of up to 2 × 10 4 . Furthermore, we implement tunable orientation attention through an intensity-responsivity matrix. In automatic driving, the feature-attention vision sensor recognizes no stopping sign with 99.52% accuracy comparable to RGB cameras while reducing data generation by 66.7%. This approach mimics this biological “sense-only-the-useful” strategy, eliminating the need for bulky optical filters and postprocessing algorithms and offering a promising pathway for the realization of compact, resource-efficient intelligent vision systems.