Integrated Photonic Processor for High‐Efficiency Megapixel RGB Image Encoding

Abstract Photonic computing demonstrates significant enhancements over conventional von Neumann architectures in artificial intelligence applications, such as machine vision, offering superior operational bandwidth and reduced energy consumption. However, current photonic computing architectures face scalability challenges when utilizing chip‐scale implementations due to inherent physical constraints and control complexities, particularly when processing high‐dimensional tensor inputs. This study presents a compact Photonic Encoding Unit (PEU) that integrates with on‐chip diffraction‐based optical connections, optimized for diverse visual tasks. The PEU employs amplitude‐phase co‐modulation to facilitate concurrent loading and processing of high‐dimensional tensors. The PEU is fabricated and experimentally validated across various single‐ and multi‐channel visual processing tasks, including grayscale image compression and denoising, megapixel color image compression with a maximum compression ratio of 4.5:1, color image classification over the CIFAR‐4 dataset with an accuracy of 70.3% and a generative image style transfer task. The results demonstrate the PEU's capability for concurrent optimization of multi‐channel vision tasks while maintaining a compact structure. This work establishes a pathway toward expanding information multiplexing dimensions in on‐chip photonic computing systems and advancing future large‐scale, high‐dimensional visual encoding systems.