Scholarlink

<p>Images taken in foggy weather suffer from degraded image quality, which can reduce the accuracy and robustness of subsequent computer vision tasks. Therefore, the restoration of clear images through image enhancement or physical models is an important technology in visual perception systems. However, the existing dehazing methods often rely on computationally intensive convolutional stacking structures, which results in high computational and memory requirements and makes deployment on resource-constrained edge platforms difficult. Therefore, developing lightweight image dehazing algorithms to meet the constraints of edge devices has become an urgent step in visual perception applications. In this paper, a lightweight image dehazing method together with physical models is proposed, which aims to achieve an efficient and effective image dehazing goal. Specifically, this paper introduces a dehazing network structure based on a Physics-Guided Neural Network (PGNN). The network structure explicitly restores key intermediate processes in the Atmospheric Scattering Model (ASM) and enhances the physical consistency of the network learning process, thereby achieving the purpose of fast and explicit recovery of hazy images. In addition, this paper proposes a Physics-Guided Loss (PGL) based on the physical model and combines it with a Visual Geometry Group (VGG) loss. This design enables the model to optimize jointly pixel-level accuracy, physical consistency, and visual perceptual quality during training. Under this loss formulation, the dehazed images are close to the ground truth at the pixel level and remain physically consistent with the atmospheric scattering principle. In order to meet the deployment needs, this paper presents a comparative performance evaluation of dehazing models under different frameworks and validates the proposed lightweight approach through systematic experiments. The experimental results demonstrate the competitive performance of PGNN in improving image quality while effectively optimizing deployment efficiency under lightweight constraints. The ablation study further supports the contribution of PGL to model performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight image dehazing via physics-guided neural networks

  • 期刊论文

  • 发表日期：

    2026年5月6日

查看全文
  • Guangyuan Liu,
  • MinPo Jung

摘要

Images taken in foggy weather suffer from degraded image quality, which can reduce the accuracy and robustness of subsequent computer vision tasks. Therefore, the restoration of clear images through image enhancement or physical models is an important technology in visual perception systems. However, the existing dehazing methods often rely on computationally intensive convolutional stacking structures, which results in high computational and memory requirements and makes deployment on resource-constrained edge platforms difficult. Therefore, developing lightweight image dehazing algorithms to meet the constraints of edge devices has become an urgent step in visual perception applications. In this paper, a lightweight image dehazing method together with physical models is proposed, which aims to achieve an efficient and effective image dehazing goal. Specifically, this paper introduces a dehazing network structure based on a Physics-Guided Neural Network (PGNN). The network structure explicitly restores key intermediate processes in the Atmospheric Scattering Model (ASM) and enhances the physical consistency of the network learning process, thereby achieving the purpose of fast and explicit recovery of hazy images. In addition, this paper proposes a Physics-Guided Loss (PGL) based on the physical model and combines it with a Visual Geometry Group (VGG) loss. This design enables the model to optimize jointly pixel-level accuracy, physical consistency, and visual perceptual quality during training. Under this loss formulation, the dehazed images are close to the ground truth at the pixel level and remain physically consistent with the atmospheric scattering principle. In order to meet the deployment needs, this paper presents a comparative performance evaluation of dehazing models under different frameworks and validates the proposed lightweight approach through systematic experiments. The experimental results demonstrate the competitive performance of PGNN in improving image quality while effectively optimizing deployment efficiency under lightweight constraints. The ablation study further supports the contribution of PGL to model performance.