Detail-enhanced dynamic masked autoencoder for non-homogeneous hazy image dehazing

Le résumé fourni par la source

Abstract Non-homogeneous hazy images are challenging to dehaze because the spatially random distribution and variable concentration of haze are difficult to determine precisely. Therefore, dehazing methods should be adaptive to the spatial characteristics of the haze. This paper proposes a detail-enhanced dynamic masked autoencoder (MAE) for non-homogeneous hazy image dehazing to address this issue. This architecture integrates three branches: the dynamic MAE reconstruction branch (DARB), the detail-enhanced capture branch (DECB), and the haze-salient attention branch (HSAB). In DARB, a dynamic masking mechanism based on HSV prior knowledge is embedded into the autoencoder of MAE. The mechanism generates masks for the MAE adaptively based on the relationship between the difference of the V and S channels in HSV and the haze concentration. This adaptive masking enables the network to perform differentiated reconstruction for regions with varying haze concentrations. Leveraging frequency-domain decomposition, the DECB and HSAB are respectively designed to enhance high-frequency detail features and capture low-frequency global haze structures, thereby enabling the network to effectively restore both fine textures and overall scene structures. By integrating three complementary branches, the proposed framework generates masks based on HSV prior information and adaptively restores clear images. Experiments on non-homogeneous haze datasets show that the proposed method effectively restores images under challenging haze conditions. On the NH-Haze dataset, it achieves improvements of 0.22 dB in PSNR and 0.0044 in SSIM over existing methods, demonstrating superior performance in both quantitative metrics and subjective visual quality.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.