Bio-inspired artificial retina with receptive fields for in-sensor multiply-and-accumulate operations

초록

Bio-inspired vision systems based on curved image sensors offer a compelling imaging hardware for mobile robots by geometrically matching the image plane to the Petzval surface of single-lens optics. However, to fully exploit the advantages of biological vision systems, it is essential not only to emulate optical structures but also to integrate sensory-level processing functions into vision hardware. Here, we propose a robotic vision system that leverages the structural advantages of the human eye (compact single-lens imaging architecture) and the functional advantages of the biological receptive fields (sensory-level information pre-processing for efficient signal transmission and downstream computation). It is enabled by a bio-inspired artificial retina composed of curved perovskite photoconductors that form artificial receptive fields (ARFs). Each ARF performs multiply-and-accumulate (MAC) operations through in-sensor computing, executing sensory-level image processing functions (e.g., edge detection). As a result, the artificial retina captures compact yet information-rich edge images without external post-processing, thereby improving the speed and energy efficiency of semantic segmentation.image

키워드