基于人类视觉系统及信号的过完备稀疏表示理论,提出一种新的多聚焦图像融合算法。首先从待融合图像中随机取块构成训练样本集,经迭代运算获取过完备字典;然后由正交匹配追踪算法完成图像块的稀疏分解;再按分解系数的显著性选择融合系数并完成图像块的重构;重构块经重新排列并取平均后获得最后的融合图像。实验结果表明:该算法继承了目前较为优秀的多尺度几何分析方法的融合效果;在噪声存在的情况下,该算法表现出较好的噪声抑制能力,随噪声方差的升高,融合图像的主观质量及客观评价指标均要好于传统方法。

A novel multi-focus image fusion algorithm based on image sparse representation theory and on the Human Visual System is introduced in this paper. For the new algorithm, the related images are partitioned into image patches. The learning for an adaptive dictionary is implemented by an iterated processing. All image patches are decomposed by the orthogonal matching pursuit (OMP) algorithm, and the coefficients with prominent properties are selected for image patch reconstruction. The reconstructed patches are realigned according to their partition order and the overlapped patches are averaged to get the fused image.The experiment results demonstrate that the proposed algorithm maintains the property of the state-of-the-art typical algorithm. In addition, with a noisy input, the proposed method can generate noiseless results and the objective criteria outperform the typical algorithm when noise variance increasing.