The compressed sensing (CS) theory has been introduced to the traditional i mage fusion procedure to improve its efficiency.However,the fusion image quality is strong ly affected by the sampling method which is used to generate the compressed coefficients.The e xisting compressed image fusion methods mainly utilize the star-shaped sampling in the Fourier fre quency domain and the radial sampling in the wavelet frequency domain.These sampling models ignor e the signal structure in the frequency field,which would readily cause a high sampling rate .Thus,it′s difficult to improve the efficiency and the performance of the current compresse d image fusion methods.In this work,the structure of the wavelet coefficients is investigated ,which indicates that the important wavelet coefficients form a quad-tree structure.Thus,a str uctured and self-adaptive sampling model is proposed for the compressed image fusion,by ta king advantage of the subtree structure which contains the important wavelet coefficients.The subtree sampling model can acquire more important coefficients because of the self-adaptive samp ling strategy. A new compressed image fusion algorithm is then proposed by combining the subtre e sampling model and the spatial recursive image reconstruction algorithm.Numerical results show that the proposed compressed image fusion algorithm can improve the image fusion performance and c omputational efficiency.