Lightweight single-image dehazing via haze-aware attention and discrete wavelet transform

Abstract Single-image dehazing is essential for safety-critical vision systems, yet existing CNN-based methods are often too heavy for edge devices or fail to adapt to spatially varying haze density. This paper introduces HazeWaveNet, a lightweight wavelet-based dehazing network that leverages discrete wavelet transform (DWT) to separate high- and low-frequency components, targeting haze-affected regions while preserving details. A haze-trend guidance mechanism, derived from the dark-channel prior, generates adaptive weight maps that modulate features through factorized convolutions and haze-aware attention modules. Extensive experiments on synthetic and real-world datasets demonstrate that HazeWaveNet achieves competitive performance with significantly lower computational requirements, making it suitable for real-time deployment on resource-constrained platforms.