Software ISP Pipeline Raw to Perceptually Correct Image (Phase 1- Ongoing)

Implementing a complete software Image Signal Processing pipeline from scratch, taking raw DNG files captured from a DJI Osmo Pocket 3 camera through every classical ISP stage to produce a perceptually accurate final image. The pipeline was built stage by stage with full visibility into each intermediate output, covering Black Level Correction (BLC) to remove sensor offset bias, Lens Shading Correction (LSC) to compensate for peripheral light falloff, Demosaicing to reconstruct full RGB from the raw Bayer pattern, White Balance to neutralize color casts from the scene illuminant, CIE XYZ color space transformation for device-independent color representation, Linear sRGB conversion, Tone Mapping to compress the high dynamic range of the raw sensor into display-suitable luminance, and finally Gamma Correction to apply the non-linear perceptual encoding the human visual system expects. Each stage was validated visually by inspecting intermediate outputs, allowing precise debugging of color accuracy and luminance response at every step. The pipeline is being extended for deployment on edge hardware targets — NVIDIA Orin Nano and Qualcomm Hexagon NPU — as Phase 2, with the goal of achieving real-time ISP processing on embedded platforms.