EdgeISPNet : Ultra-Lightweight neural ISP for High-Resolution RAW-to-RGB High Altitude Imaging Systems

Kurzfassung

Onboard computer vision pipelines for High-Altitude Platform System (HAPS) require real-time, high-fidelity RAW-to-RGB conversion under extreme power and weight constraints—a scenario not addressed by current neural ISP methods, which focus on smartphone sensors and near-ground illumination. We identify three critical gaps:(1) no neural ISP architecture is designed for stratospheric imaging, (2) no edge-deployable model supports 150 MP tiled inference with global consistency, and (3) no benchmark dataset exists for this domain. We address all three by introducing EdgeISPNet, an ultra-lightweight,re-parameterizable network with ∼28 K inference-time parameters that maps 16-bit Bayer RAW to RGB using only standard operators and compiles natively on any edge accelerator. Our novel Global Scene Representation Module (GSRM) ensures seamless tonal consistency across tiles. On a Jetson Orin Nano, EdgeISPNet processes a full 150 MP image in 4.16 s end-to-end—a 3× throughput improvement over the vendor SDK. Additionally, we make available the first paired aerial RAW-to-RGB dataset consisting of 130 full-resolution 150 MP images captured from a manned aircraft under diverse atmospheric conditions.