李瑞俊,高霞.惩罚KL散度耦合迭代分布加权的图像重构算法[J].包装工程,2015,36(3):107-112,146.
LI Rui-jun,GAO Xia.Image Reconstruction Algorithm Based on Penalizing KL Divergence Coupled Iteration Distribution Reweighting[J].Packaging Engineering,2015,36(3):107-112,146. |
| 惩罚KL散度耦合迭代分布加权的图像重构算法 |
| Image Reconstruction Algorithm Based on Penalizing KL Divergence Coupled Iteration Distribution Reweighting |
| 投稿时间：2014-07-16 修订日期：2015-02-10 |
| DOI： |
| 中文关键词: 图像重构 惩罚KL散度 迭代分布重新加权 MAP估算 图像梯度分布 用户响应 |
| 英文关键词: image reconstruction penalizing KL divergence iteration distribution reweighting MAP estimation image gradient distribution user response |
| 基金项目:内蒙古自治区自然科学基金 （2012MS0931） |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 利用MAP（Maximizing A Posteriori）估算技术能够较好地重构退化图像， 有效降低复原图像的伪影， 但MAP估算是惩罚图像的非零梯度， 且其图像先验估算忽略了退化图像自身纹理， 易导致重构图像过渡平滑， 产生突变阶跃边缘和丢失图像中频纹理信息。设计了惩罚KL散度耦合迭代分布重新加权的图像重构算法。方法 基于退化图像像素， 构建图像参考梯度分布计算模型， 以估算图像先验； 引入KL（Kullback-Leibler）散度， 联合MAP估算技术， 惩罚经验分布与参考分布之间的梯度； 设计迭代分布重新加权算法， 最小化成本函数， 优化经验梯度分布， 使其更接近参考分布； 基于 HVS（Human Visual System）， 构造了转导对比度失真率模型。最后， 利用 Amazon Mechanical Turk数据集， 对提出的算法进行用户响应研究。结果 仿真实验结果表明， 与当前基于MAP估算技术的图像重构机制相比， 在图像退化程度严重时， 提出的算法具有更好的用户响应， 且具有更高的重构精度， 复原图像的纹理细节清晰可见。结论 提出的算法具有更高的重构质量， 用户响应良好。 |
| 英文摘要: |
| The MAP(maximizing a posteriori) estimation can be use to reconstruct degraded images and effectively reduce the artifacts of the restored images; but the MAP estimation is non-zero gradient of penalty images, and the image priori estimation ignores the own texture of the degraded images, resulting in overly smooth restored images, with abrupt step edges and a loss of mid-frequency texture information. Therefore, the image reconstruction algorithm based on MAP estimator coupled iteration distribution reweighting was proposed. The image conference gradient distribution calculation model was constructed based on the pixel of degraded images, to estimate the image priori. The KL divergence was introduced, in combination with MAP estimation, to penalize the gradients between the empirical and reference distributions. And iterative distribution reweighting algorithm was designed to minimize the cost function and optimize the empirical gradient distribution for improving the convergence precision, and making the empirical gradient distribution closer to reference distributions. Transduction contrast distortion model was constructed based on Human Visual System. Finally, user response study was conducted for the proposed algorithm using Amazon Mechanical Turk. The reconstruction quality of the proposed algorithm was high, and the user response was good. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
渝公网安备 50010702501716号