High dynamic camera design for night use combined with AI-ISP

Abstract In recent years, Artificial Intelligence Image Signal Processor (AI-ISP) have garnered significant attention in the field of low-light image restoration. Most existing low-light restoration methods primarily focus on enhancing dark regions, often resulting in over-saturation in brightly lit areas of the scene. Additionally, many models suffer from excessive parameter counts, making them unsuitable for practical deployment. To address these challenges, this paper proposes a lightweight, end-to-end AI-ISP network. The network comprises a base module for integrated image feature extraction and a weight modulation module designed to handle high-dynamic regions in images, enabling both low-light restoration and glare suppression. Furthermore, a specialized dataset featuring high-dynamic image effects has been constructed to support the training of the proposed AI-ISP network, enhancing its performance for nighttime applications. In addition, this study discusses the model pruning, quantization and parallel acceleration strategy combined with the process pool to optimize the computational efficiency. The proposed algorithm is successfully deployed on the RK3588 embedded platform to develop a high-dynamic night vision camera. Experimental results demonstrate that the AI-ISP-powered night vision camera can recover low-light details from RAW images with a resolution of 1 × 768 × 1024 at ~24 frames per second, while simultaneously delivering superior glare suppression and high-dynamic performance.