KNOWLEDGE-DISTILLED VARIATIONAL BAYESIAN FRAMEWORK FOR EFFICIENT LARGE-SCALE IMAGE DEHAZING

Abstract

Real-time processing of large-scale image databases with state-of-the-art dehazing methods presents a significant computational challenge: methods that achieve superior generalization typically require substantial inference time, limiting their deployment in real-time applications. High-performing methods employ complex multi-scale processing and deep architectures, typically achievin less than 5 FPS on high-resolution images. Building on the preliminary multi-scale variational Bayesian framework [1, 2], which achieves strong synthetic-to-real generalization, this paper proposes knowledge distillation to transfer the generalization capabilities of high-performance models to a lightweight Vision Transformer-based student network. The student leverages patch-based processing and reduced architectural complexity to achieve over $150\times$ speedup, while maintaining competitive performance through a theoretically-grounded distillation framework integrated into the variational Bayesian objective. Additionally, the atmospheric scattering model is extended to estimate space-variant atmospheric light, improving performance on varying haze regions. Trained solely on synthetic Haze4K data, the proposed method stays competitive on synthetic-to-real generalization and downstream object detection (on the augmented KITTI dataset) tasks, while achieving superior inference speed for large-scale real-world applications.