Abstract: Aiming at the problems of poor image quality, lack of polarization information, and inadequate target texture details in current  infrared polarization image fusion,  we proposed an  infrared polarization image fusion method based on composite domain multi-scale decomposition. Firstly, in the spatial domain, a two-scale decomposition of the source image was performed by using a bootstrap filter to obtain the detail and base layers, in the frequency domain, a multi-scale multi-directional decomposition of the base layer image was performed by using a non-subsampled shear-wave transform to obtain the low-frequency sub-band image and high-frequency  sub-band image.  Secondly, the principal component analysis-adaptive pulse coupled neural network fusion rule was used for  high-frequency sub-band,  an improved convolutional sparse representation was used for coefficient merging for the low-frequency sub-bands, and  local energy weighting and selective fusion rules based on pixel similarity were used for detail layed fusion. Finally, the fused image was reconstructed by using an inverse transformation in the composite domain. Experimental results show  that the proposed method outperforms other comparative fusion methods in  subjective visual performance and eight objective evaluation metrics,  indicating that the method has many advantages in infrared polarization image fusion and can effectively enhance the quality of fused images.

Key words: infrared polarization image fusion, non-subsampled shear-wave transform, adaptive pulse coupled neural network, convolutional sparse representation