| 基于对比学习的单幅图像去雾算法 |
| A single image defogging algorithm based on contrast learning |
| |
| DOI： |
| 中文关键词: 对比学习；特征提取；特征金字塔；特征融合；图像去雾 |
| 英文关键词:contrast learning； feature extraction； feature pyramid； feature fusion； image defogging |
| 基金项目:国家重点研发计划（2019YFB2103003）资助项目 |
| 作者 | 单位 | | 何 涛 | 南京邮电大学 电子与光学工程学院、柔性电子（未来技术）学院，江苏 南京 210023 | | 许广峰 | 南京邮电大学 电子与光学工程学院、柔性电子（未来技术）学院，江苏 南京 210023 | | 徐 鹤 | 南京邮电大学 计算机学院、软件学院、网络空间安全学院，江苏 南京 210023；江苏省无线传感网高新技术研究重点实验室，江苏 南京 210023 | | 杨 忆 | 南京邮电大学 电子与光学工程学院、柔性电子（未来技术）学院，江苏 南京 210023 |
|
| 摘要点击次数: 68 |
| 全文下载次数: 21 |
| 中文摘要: |
| 现有的图像去雾算法，只采用了清晰图像来指导去雾网络的训练，而没有利用模糊图像，从而造成去雾不彻底，细节信息不完整的问题。为此提出了一种对比正则化的方法，利用模糊图像和清晰图像共同指导去雾网络的训练。对比正则化保证恢复后的图像信息向清晰图像方向靠近，远离模糊图像的方向。此外提出一种新的金字塔通道的特征自适应融合网络。该网络包含3个部分：三尺度特征提取网络、特征自适应混合模块(PCFM)和图像重建模块。三尺度特征提取模块同时捕捉不同尺度特征。金字塔结构和特征自适应融合操作，有效地提取相互依赖地特征，并以金字塔的方式有选择性地聚集更重要的特征。图像重建模块用于重建特征，恢复清晰的图像。实验结果表明，与现有的经典去雾算法相比，客观评价指标：峰值信噪比(PSNR)和结构相似性(SSIM)都得到了提升，并改善了去雾不彻底和颜色失真的现象。 |
| 英文摘要: |
| The existing image defogging algorithms only use clear images, instead of fuzzy ones, to guide the training of defogging networks, and result in incomplete defogging and incomplete details. We propose a method of contrast regularization to guide the training of defogging networks with both fuzzy images and clear images. Contrast regularization ensures that the restored image information is close to the clear image and far away from the blurred image. In addition, we propose a new pyramid channel feature adaptive fusion network. The network consists of three parts: a three-scale feature extraction network, a feature adaptive mixing module (PCFM) and an image reconstruction module. The three scale feature extraction network captures different scale features at the same time. The pyramid structure and the feature adaptive fusion operation can effectively extract interdependent features, and selectively aggregate more important features in the form of pyramid. The image reconstruction module is used to reconstruct features and restore clear images. Experimental results show that the objective evaluation indexes of the proposed algorithm, including the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM), have been improved compared with those of the existing classical defogging algorithms, and the problems of incomplete defogging and color distortion have been decreased by the proposed algorithm. |
| 查看全文 查看/发表评论 下载PDF阅读器 |