Multi-Branch Non-Homogeneous Image Dehazing via Concentration Partitioning and Image Fusion

2026-05-05 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

Researchers from Zhejiang Gongshang University and Tiangong University propose CPIFNet, a novel multi-branch deep neural network framework designed to address non-homogeneous image dehazing. Existing methods often struggle with spatially varying haze concentrations, leading to over- or under-enhancement. CPIFNet tackles this by decomposing a non-homogeneous hazy image into multiple local regions, each treated as approximately homogeneous. The framework consists of a two-stage architecture: an Image Enhancement Network (IENet) stage with multiple branches independently trained on homogeneous haze datasets of different concentration levels, and an Image Fusion Network (IFNet) stage that intelligently aggregates advantageous regions from the IENet outputs. CPIFNet is supervised by a comprehensive loss function including reconstruction, perceptual, structural, and color losses. Experiments on synthetic benchmarks (FiveK-Haze, SOTS-indoor, SOTS-outdoor) and real-world datasets demonstrate state-of-the-art performance, with PSNR improvements of 5.29 dB on FiveK-Haze, 2.52 dB on SOTS-indoor, and 2.83 dB on SOTS-outdoor over the second-best methods.

Key takeaway

For research scientists developing image restoration models, CPIFNet's multi-branch, concentration-partitioning approach offers a robust strategy for non-homogeneous dehazing. You should consider adopting a similar decomposition and fusion architecture, particularly for tasks involving spatially varying degradation, to achieve superior performance and visual quality compared to single-pathway networks. Evaluate the optimal number of concentration partitions for your specific dataset to balance specialization and data sufficiency.

Key insights

Decomposing non-homogeneous haze into homogeneous sub-problems enables targeted, multi-branch image enhancement and fusion.

Principles

• Haze concentration varies spatially within images.
• Extreme haze data can degrade model training.
• Stacking fusion outperforms weighted fusion.

Method

CPIFNet uses multi-branch IENets, each trained on specific haze concentrations, followed by an IFNet that aggregates deep features from these branches to produce a unified dehazed image.

In practice

• Partition training data by haze concentration.
• Exclude extremely dense haze data from training.
• Use comprehensive loss functions for supervision.

Topics

• Non-Homogeneous Image Dehazing
• Multi-Branch Neural Networks
• Concentration Partitioning
• Image Fusion
• CPIFNet

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• Towards UAV Image Dehazing: A UAV Atmospheric Scattering Model, Benchmark, and Geometry-Aware Deep Unfolding Network — A new UAV atmospheric scattering model and deep unfolding network effectively dehaze UAV images by accounting for non-uniform haze geometry.
• SWNet: A Cross-Spectral Network for Camouflaged Weed Detection — Cross-spectral fusion and edge refinement are crucial for accurate camouflaged weed detection in agriculture.
• The Third Challenge on Image Denoising at NTIRE 2026: Methods and Results — The NTIRE 2026 challenge benchmarks unconstrained neural architectures for high-noise image denoising, prioritizing PSNR performance.
• IG-Diff: Complex Night Scene Restoration with Illumination-Guided Diffusion Model — Illumination-guided diffusion models effectively restore complex night scenes by integrating illumination priors and synthetic data.
• PRISM: Synergizing Vision Foundation Models via Self-organized Expert Specialization — PRISM synergizes Vision Foundation Models through a dual-stream Mixture-of-Experts framework with self-organized expert specialization.
• NTIRE 2026: A Multi-Track Study in Image Restoration — NTIRE workshops benchmark image restoration models against real-world degradations using rigorous evaluation protocols.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.