Streaming Fixed-Point Datapath for Retinex-Inspired Low-Light Image Enhancement: Open-Source Synthesis and Timing Analysis Towards Agricultural Edge Vision

Low-light conditions can degrade agricultural vision imagery, while edge camera nodes may have limited computational resources. This study developed and characterized a streaming, fixed-point FPGA datapath for Retinex-inspired low-light image enhancement and evaluated the underlying algorithm in software. The register-transfer-level (RTL) design combines maximum-channel illumination estimation, edge-aware refinement with piecewise linear Gaussian weighting, table-based reflectance recovery and gamma remapping, and batch-normalization-style output normalization. It processes one pixel per clock cycle without a frame buffer. Functional simulation matched a bit-accurate arithmetic model at every compared pixel. Open-source synthesis, placement, and routing targeting an AMD Xilinx Artix-7 FPGA yielded 5901 LUTs and met a 100 MHz timing constraint. Image quality was evaluated only in software. On 15 LOL-v1 test pairs, a floating-point reference model differing from the RTL in two processing steps improved mean PSNR from 7.77 to 15.03 dB; an evaluation using the RTL equations achieved 9.35 dB, increasing to 16.64 dB after reselecting a normalization constant. The findings indicate a compact streaming realization of the pipeline, while highlighting the sensitivity of image quality to normalization. The design was not programmed or tested on a physical FPGA board, and real agricultural low-light images were not evaluated.