A Robust Real-Time Dehazing Framework Based on a Dimensional-Adaptive Dark Channel Prior for Intelligent Visual Sensing

Visual perception systems in autonomous robotics are severely degraded by atmospheric haze, which reduces contrast and obscures structural details required for navigation. While modern Deep Learning models have achieved significant progress in dehazing, their heavy reliance on matrix operations renders them computationally prohibitive for real-time deployment on standard Central Processing Units (CPUs). To address this bottleneck, this paper proposes the Dimensional-Adaptive Patchsize Dark Channel Prior (DAP-DCP), a lightweight framework designed for high-frequency embedded visual sensing. The core innovation is the DAP algorithm, which dynamically calibrates the analysis kernel to 4% of the image’s minimum dimension, theoretically ensuring scale-invariant robustness against halo artifacts without the latency of segmentation networks. This adaptive prior is integrated into a coarse-to-fine optimization pipeline, featuring a hybrid sky-priority atmospheric light estimation and a Guided Filter-based upsampling strategy to preserve high-frequency details. Experimental benchmarking on the SOTS Indoor dataset demonstrates that DAP-DCP achieves a Structural Similarity Index (SSIM) of 0.9319, outperforming the lightweight deep learning baseline AOD-Net (0.9092). Crucially, the framework operates at 0.0521 seconds per frame (19 FPS) on a CPU environment, representing a 12x speedup over AOD-Net. Furthermore, validation on real-world hazy scenes using the YOLOv11 detector confirms that the proposed restoration increases object detection recall by 14.3% and improves global confidence scores by 2.9%. These results establish DAP-DCP as a superior engineering solution for real-time robotic vision in resource-constrained environments.