Real-time FPGA video dehazing using a lightweight improved dark channel prior

Abstract

Hazy environments degrade the accuracy of traffic surveillance systems owing to blurred frames and loss of image detail. To reduce hardware resource consumption while mitigating the sky-region colour distortion of the dark channel prior (DCP) dehazing algorithm, three lightweight improvements are proposed: a sky segmentation method based on the minimum channel map and four-directional Sobel operator, a simplified atmospheric light estimation that eliminates sorting overhead by exploiting the near-uniform RGB property of sky regions, and an adaptive transmittance compensation mechanism driven by pixel-wise minimum-channel differences. A pure-FPGA pipelined architecture exploits inter-frame similarity and fixed-point quantisation to further minimise resource usage, integrating OV5640 acquisition with HDMI display. Experimental results demonstrate that the system processes 1280×720 video at 30 fps while consuming only 1.19% DSP and 10.72% FF resources on the PGL50H device.