Self-Supervised Domain Adaptation in Crowd Counting

Atmospheric haze degenerates image visibility or visual quality. Unfortunately, characterizing scene depth and scene radiance uncertainties to precisely recover the transmission still remains challenging. This paper proposes a new structure-aware transmission recovery strategy for single image dehazing. Specifically, such a new strategy is a coarse-to-refined dehazing approach that integrates multiscale convolutional neural networks and structure-aware retinex modeling into filtering-based fusion to accurately recover the transmission map. The proposed method was evaluated on benchmark data and compared with currently available dehazing approaches by quantifying colorfulness, contrast, and structural fidelity information of dehazed photographs. The experimental results demonstrate that the robust structure-preserving dehazing method outperforms other approaches, with significantly improving the average visual quality or content of the colorfulness, contrast, structural sharpness, and visibility from (0.63, 0.72, 0.15, 0.62) to (0.81, 0.84, 0.18, 0.78).