Boosting Image Dehazing via Elaborate Integration of Complementary Dependencies

Xiaozhe Zhang, Fengying Xie, Haidong Ding, Linpeng Pan, Zhenwei Shi

Keywords: Varying Range Dependencies, Haze Density Map Estimation, Image Dehazing

TL;DR: We are the first to demonstrate that effective aggregation of varying range dependencies enables robust haze removal, guided by our first-of-its-kind joint estimation of the haze density map and semantic map.

Abstract: Haze removal seeks to restore clear images from hazy inputs. Previous research demonstrates that short-range dependencies are effective for preserving local details, while long-range dependencies capture global context. Because both are essential to dehazing and complement each other, many approaches explicitly integrate them within dual-stream frameworks. However, the trustworthy aggregation of these dependencies remains underexplored. In this paper, to optimize the contributions of dependencies at varying ranges, we first conduct comprehensive quantitative and qualitative experiments to identify the key influencing factors. Our findings indicate that an effective aggregation strategy should jointly consider haze density and semantic information. Building on these insights, we introduce a CLIP-enhanced Dual-Path Aggregator for the class of dual-stream dehazing methods. This module first employs a shared backbone to generate fine-grained haze density and semantic maps in a computationally efficient manner, and then uses them to instruct the integration process. Extensive experiments show that the proposed aggregator significantly improves the performance of existing dual-stream methods, and our custom-built model, DehazeMatic, achieves state-of-the-art results across multiple benchmarks. As an additional contribution, we also address, for the first time, the challenge of accurately estimating haze density maps.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6545