基于支持向量聚类的多聚焦图像融合算法

Exploiting SVC Algorithm for Multifocus Image Fusion

从无监督机器学习角度提出了一种基于SVC(support vector clustering)的图像融合规则,解决了基于SVM(support vector machine)的融合规则在处理多聚焦图像融合问题时所引起的区域混叠与非平滑过渡问题,进一步提高了融合图像的质量.使用非降采样离散小波变换对源图像进行多分辨率分解,基于网格提取源图像的特征.图像特征集合作为SVC的输入数据集,聚类结果最终由区域鉴别算法分配到两个区域:互补信息区域和冗余信息区域,并分别采用选择法和加权平均法生成融合图像的多分辨率表示,通过对这一多分辨率表示进行小波逆变换重构融合图像.详细研究了SVC的参数q与融合效果的评价参数RMSE之间的关系.理论分析及实验结果均表明,SVC用于图像融合问题是合适的,而且比较实验显示,基于SVC的融合规则优于基于SVM的融合规则.

This paper proposes a SVC （support vector clustering） based fusion rule according to unsupervised learning strategy. By employing the rule in multifocus image fusion applications, it solves the problems of region overlapping and abrupt transition brought about by the SVM （support vector machine） based fusion rule. The quality of the fused image is further enhanced. The undecimated discrete wavelet transform is applied to source images for multiresolution decomposition. Image feature data is extracted by means of grid, and it is then fed into the SVC algorithm which will generate distinct clusters. These resultant clusters are further processed by the domain discrimination algorithm and eventually distributed to two separate domains defined as complementary domain and redundant domain, in which choose-max method and weighted average method are used respectively to produce multiresolution representation of the fused image, Finally, the fused image is reconstructed by performing the corresponding inverse wavelet transform. The relation between the parameter q of SVC algorithm and the parameter RMSE used to evaluate the fused image is studied in detail. It is indicated by theoretical analysis and experimental results that SVC is appropriate for image fusion. Moreover, comparative studies show that the proposed SVC based fusion rule outperforms the existing SVM based ones.