提出了一种非抽样双树复小波变换(UDT-CWT)与基于块主元旋转的非负矩阵分解(BPP-NMF)相结合的多聚焦图像融合算法。利用UDT-CWT具有完美的平移不变性及良好的方向选择性, 首先对图像进行多尺度、多方向分解并得到低频子带和高频子带系数; 然后对低频子带系数采用块主元旋转的非负矩阵分解的融合策略, 高频系数则选用高斯加权区域能量与区域标准差一致性选择的融合准则。最后对融合后的系数进行UDT-CWT逆变换得到重构图像。选用多组多聚焦图像进行融合并对融合结果进行主观视觉、客观方面的评价。试验结果表明, 该融合算法不仅具有良好的视觉效果, 同时在客观评价指标也优于一般的融合策略, 验证了该算法的有效性。

A multi-focus image fusion algorithm based on undecimated dual-tree complex wavelet transform(UDT-CWT) domain and block principal pivoting method for nonnegative matrix factorization(BPP-NMF) is proposed. Firstly, combined with its shift invariance and good directional selectivity, the source images are decomposed to sub-band images with multi-scales and multi-directions by UDT-CWT, thus the low frequency sub-band coefficients and high frequency sub-band coefficients are obtained. The fusion strategy of low frequency sub-band coefficients was based on the method of nonnegative matrix factorization which adopted the block principal pivoting method; while for high frequency sub-band, a fusion rule adopting the consistency selection of Gaussian-weighted regional energy and area standard deviation is presented. Finally, the fused low and high frequency coefficients are reconstructed to image by undecimated dual-tree complex wavelet inverse transform. Fusion result is done with several sets of different multi-focus images, both subjective visual evaluation and objective performance assessments of these results are implemented. Results indicate that the fusion algorithm not only has good subjective visual effects, but also better than other general fusion strategies in the objective evaluation; hence the algorithm is effective.