Scholarlink

<p>Existing deep learning-based dehazing methods are predominantly result-oriented and lack process guidance while they fail to fully exploit the implicit features of clear images. To address the limitations this paper proposes an encoder-decoder network structure based on a teacher-student model, called KG-Net. Specifically, both the teacher network and the student network adopt the same encoder-decoder architecture, comprising a Shallow-Deep Feature Fusion module (SFF), a Feature Perception module (FP), and a Dynamic Feature Enhancement module (DFE). The SFF module adaptively fuses features between the downsampling and upsampling layers, preserving the shallow detail features and the deep semantic features. The FP module enhances the network’s ability to capture richer feature representations, while the DFE module strengthens spatial feature learning. A loss function composed of reconstruction and perceptual losses is employed to pretrain the teacher network on clear images, enabling the extraction of implicit clean image features. Subsequently, the student network is trained using a combination of reconstruction, contrastive, and knowledge transfer losses, where knowledge transfer guided the dehazing process by leveraging the implicit features from the teacher network. Experimental results demonstrate that the proposed method outperforms existing approaches on benchmark datasets, including SOTS, BeDDE, and I/O-HAZE.The source code will be available at <a href="https://github.com/JNmcZhw/KG-Net">https://github.com/JNmcZhw/KG-Net</a>.</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

KG-Net: An effective knowledge-guided framework for single image dehazing via feature enhanced representation

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Hongwei Zhang

摘要

Existing deep learning-based dehazing methods are predominantly result-oriented and lack process guidance while they fail to fully exploit the implicit features of clear images. To address the limitations this paper proposes an encoder-decoder network structure based on a teacher-student model, called KG-Net. Specifically, both the teacher network and the student network adopt the same encoder-decoder architecture, comprising a Shallow-Deep Feature Fusion module (SFF), a Feature Perception module (FP), and a Dynamic Feature Enhancement module (DFE). The SFF module adaptively fuses features between the downsampling and upsampling layers, preserving the shallow detail features and the deep semantic features. The FP module enhances the network’s ability to capture richer feature representations, while the DFE module strengthens spatial feature learning. A loss function composed of reconstruction and perceptual losses is employed to pretrain the teacher network on clear images, enabling the extraction of implicit clean image features. Subsequently, the student network is trained using a combination of reconstruction, contrastive, and knowledge transfer losses, where knowledge transfer guided the dehazing process by leveraging the implicit features from the teacher network. Experimental results demonstrate that the proposed method outperforms existing approaches on benchmark datasets, including SOTS, BeDDE, and I/O-HAZE.The source code will be available at https://github.com/JNmcZhw/KG-Net.