基于二代曲波变换和PCNN的高光谱图像融合

Fusion of hyperspectral images based on second generation curvelet transform and pulse-coupled neural networks

为了解决高光谱图像的高数据维给后续图像分析和处理带来的困难,提出了一种基于二代曲波变换和脉冲耦合神经网络(PCNN)的融合新算法.利用各波段数据间的局部相关性将整个数据空间划分为若干个相关性较强的独立子空间,在对子空间内的各波段光谱图像进行曲波多分辨率分解的基础上,分别依据各波段图像所含有的信息量对曲波粗尺度系数进行加权融合和利用PCNN的全局耦合特性与脉冲同步特性对细尺度系数进行智能选取,最后由曲波逆变换得到各子空间的融合图像.AVIR IS数据融合实验表明,该算法能有效地实现高光谱数据维数减少和特征提取,相比于提升小波融合算法、主成分变换算法和基于典型融合准则的曲波融合算法,其所提取的图像特征在高光谱异常检测时能得到更多的真实目标.

A new fusion algorithm based on second generation curvelet transform and pulse-coupled neural networks （PCNN） was developed in order to solve problems in image processing caused by high dimensions in hyperspectral images. First, according to local correlation between neighboring bands, the whole data space was divided into several independent subspaces with relatively strong correlation; a multi-resolution curvelet decomposition was then performed to images in various bands of subspace. After these steps the coarse scale coefficients from curvelet transform were made a weighted fusion based on the entropy of images in different bands; and the fine scale coefficients were selected intelligently using the global coupling and pulse synchronization characteristics of PCNN. Finally, the fused coefficients were reconstructed to obtain the fusion image of every subspace by an inverse curvelet transform. Fusion experiments on AVIRIS data showed that the proposed method, along with typical fusion rules for target de- tection in hyperspectral images, is very effective in reduction of data dimension and feature extraction. Compared with lift-wavelet fusion algorithm, principal components transform algorithm and the curvelet fusion algorithm based on typical fusion criterion, the image features extracted by the proposed method can get more real objects in hyperspectral abnormal detection.