Scaling Up Single Image Dehazing Algorithm by Cross-Data Vision Alignment for Richer Representation Learning and Beyond

TL;DR

This work tackles the domain-gap challenge in large-scale image dehazing by introducing cross-data vision alignment that combines external augmentation to harmonize distributions across hazy datasets with internal weak-to-strong self-supervised augmentation to enrich local details. The approach integrates a Swin-transformer–based dehazing network with a dual loss that separately optimizes external reconstruction and internal detail consistency, enabling richer representations. Empirical results on NID and RSID show substantial improvements over state-of-the-art baselines in PSNR, SSIM, MSE, and FSIM, along with qualitative gains in texture and color fidelity. The framework advances robust, generalizable dehazing capable of scaling across datasets, albeit with higher training time, pointing to future work on efficiency.

Abstract

In recent years, deep neural networks tasks have increasingly relied on high-quality image inputs. With the development of high-resolution representation learning, the task of image dehazing has received significant attention. Previously, many methods collect diverse image data for large-scale training to boost the performance on a target scene. Ignoring the domain gap between different data, former de-hazing methods simply adopt multiple datasets for explicit large-scale training, which often makes the methods themselves be violated. To address this problem, we propose a novel method of cross-data vision alignment for richer representation learning to improve the existing dehazing methodology. Specifically, we call for the internal- and external knowledge should be further adapted with a self-supervised manner to fill up the domain gap. By using cross-data external alignment, the datasets inherit samples from different domains that are firmly aligned, making the model learn more robust and generalizable features. By using the internal augmentation method, the model can fully exploit local information within the images, and then obtaining more image details. To demonstrate the effectiveness of our proposed method, we conduct training on the Natural Image Dataset (NID). Experimental results show that our method clearly resolves the domain gap in different dehazing datasets and presents a new pipeline for large-scale training in the dehazing task. Our approach significantly outperforms other advanced methods in dehazing and produces dehazed images that are closest to real haze-free images.

Figures (12)

• Figure 1: Comparison with state-of-the-arts in terms of flowchart and characteristics. With the rapid development of Transformer dehazeformerselfswintransformer30, the demand for data has grown at the same time. However, the simple combination of multiple datasets achieves unsatisfactory results. To address this issue, we call for the internal- and external knowledge should be further augmented by vision alignment and self-supervised learning to perform an effective large-scale training.
• Figure 2: Illustration of the proposed integrated framework for External-Augmentor with Auxiliary Datasets (External-Aug. with AD.) and Internal-Augmentor with Self-supervised Learning (Internal-Aug. with SSL.). In our method, we perform external augmentation on the target dataset to increase its diversity. The external augmentor makes the auxiliary dataset be aligned with the target dataset in the color channels. And gamma correction mainly affects the brightness and contrast of the images. Then, we apply the self-supervised learning method to augment the data internally. Specifically, the internal augmentor emphasizes attention to fine-grained details in images. By simultaneously augmenting the data externally and internally, our model achieves higher performance and robustness.
• Figure 3: A demonstration of domain gap between different dahazing datasets. We perform gamma correction on the RSID based on the RGB values of the NID. It can be observed that before correction, there is a significant difference between the distributions of the datasets in the RGB channels. After external- augmentation, the distributions of RSID images are closely aligned with those of NID images in the RGB channels.
• Figure 4: Illustration of external augmentation. The RSID images become closer to the NID dataset by external augmentor.
• Figure 5: Dehazed images obtained with and without internal data augmentor.