Guided-Filter Transmission Refinement for Dark Channel Prior-Based Single Image Dehazing

Haze scatters and attenuates light before it reaches an imaging sensor, degrading the contrast and colour fidelity that outdoor vision systems such as autonomous driving, surveillance, and remote sensing depend on. The classical Dark Channel Prior (DCP) recovers a haze-free image by estimating a per-pixel transmission map, but its original box-filtered transmission blurs across depth discontinuities and introduces halo artefacts near edges. This paper implements and quantitatively evaluates a lightweight fix: replacing the box-filtered transmission with an edge-aware guided-filter refinement, and separately testing a CLAHE-based local-contrast post-enhancement stage. Since real hazy images lack paired ground truth, the three variants -- box-filtered baseline, guided-filter-refined, and guided-filter-plus-CLAHE -- are evaluated on a physically synthesised five-image benchmark with known ground truth, using PSNR and SSIM. Guided-filter refinement improves PSNR on three of five images and SSIM on four of five, giving a small net average gain over the baseline, whereas CLAHE post-enhancement raises perceptual contrast but reduces average fidelity to ground truth. These results quantify a concrete, reproducible edge-aware improvement to DCP and expose a fidelity-versus-perceptual-quality trade-off relevant to choosing a post-processing strategy for latency-constrained applications such as driver-assistance and surveillance systems.