| 高雯婷,廖苗,荣佳丽,郭娟秀.基于误差反馈和雾霾感知的图像去雾[J].电子测量与仪器学报,2025,39(11):214-223 |
| 基于误差反馈和雾霾感知的图像去雾 |
| Image dehazing based on error feedback and haze aware |
| |
| DOI： |
| 中文关键词: 图像去雾 误差反馈 雾霾感知 注意力机制 |
| 英文关键词:image dehazing error feedback haze aware attention mechanism |
| 基金项目:湖南省自然科学基金(2025JJ50363)、湖南省教育厅科研项目(24A0356)、湖南省科技创新计划(2024RC3216)项目资助 |
|
|
| Author | Institution |
| Gao Wenting | School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan 411100, China |
| Liao Miao | School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan 411100, China |
| Rong Jiali | School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan 411100, China |
| Guo Juanxiu | School of Computer Science and Engineering, Hunan University of Science and Technology, Xiangtan 411100, China |
|
| 摘要点击次数: 336 |
| 全文下载次数: 345 |
| 中文摘要: |
| 雾霾环境下拍摄的图像通常受到对比度降低、细节退化或色彩失真等多种因素的干扰，严重影响视觉体验和后续高级视觉任务的准确性。为了有效去除图像中的雾霾,提出了一种基于误差反馈的多尺度密集残差去雾网络(multi-scale dense residual dehazing network, MDRD-Net)。该网络在编、解码路径中对称引入误差反馈模块(error feedback module, EFM),以补偿下采样导致的信息丢失,并在误差反馈模块之间引入密集连接,加强非邻层级间的信息交互。为了让网络重点关注图像中的浓雾和细节区域,网络在特征提取阶段级联了多个雾霾感知模块(haze aware module, HAM)。此外,网络在跳跃连接中引入注意力机制,实现对编、解码端特征的自适应融合,克服深、浅层特征之间存在的语义鸿沟。在RESIDE公开数据集上进行了广泛实验,结果表明,提出方法可有效去除图像中的雾霾干扰,获得色彩真实、对比度高、细节丰富的高质量无雾图像,在定量和定性分析中均显著优于多种现有先进方法。 |
| 英文摘要: |
| Images captured in haze are often affected by contrast reduction, detail degradation, or color distortion, which significantly impair visual quality and affect the performance of high-level vision tasks. To effectively remove the haze from images, a multi-scale dense residual dehazing network (MDRD-Net) based on error feedback is proposed. In this network, error feedback modules (EFM) are symmetrically introduced in the encoding and decoding paths to compensate for the information loss caused by downsampling. Dense connections are introduced between error feedback modules to enhance information interaction between non-adjacent layers. To make the network focus on regions with thick haze and rich details, multiple haze aware modules (HAM) are cascaded in the feature extraction stage. Additionally, an attention mechanism is introduced in the skip connections to adaptively fuse the features from the encoder and decoder to overcome the semantic gap between deep and shallow features. Extensive experiments on the RESIDE public dataset demonstrate that the proposed method can effectively remove the haze interference and obtain clear images with true colors, high contrast, and rich details. The results, both quantitatively and qualitatively, show a significant improvement over those of many existing state-of-the-art methods. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
