Research On Image Fusion Based On Tensor

As an emerging research field, image fusion has made some achievements in the last 30 years and has been widely applied in military, medicine and remote sensing. However, there are still many theoretical and technical problems to be solved. Especially, domestic research on image fusion starts late and is relatively backward. Thus, it is urgent to carry out in-depth research on image fusion.Tensor-based information processing methods are more suitable for representing high dimensional data and extracting relevant information than vector-based and matrix-based methods, thus receiving lots of attention. Especially, Higher Order Singular Value Decomposition(HOSVD) is an efficient data-driven decomposition technique, and the process of decomposition does not require setting parameters and thresholds. As another efficient image analysis tools, the structure tensor can efficiently analyse image structure and estimate image direction.On the basis of learning traditional image fusion methods and tensor, we have made a further research on image fusion based on tensor in the paper. It mainly relates to the application of HOSVD and structure tensor in image fusion. The main contributions of this paper are as follows:(1) Two multi-focus image fusion methods based on HOSVD are proposed. In order to solve the uncertainties of the source images’ contributions to the fused image, a multi-focus image fusion method based on HOSVD and fuzzy reasoning is proposed. This method combines the advantages of HOSVD and fuzzy reasoning. The sliding window technique is used to divide the image into patches and HOSVD is employed to extract the decomposition coefficients of sub-tensors. And fuzzy reasoning rule is designed based on average energy, regional energy and match degree. For the source image may exist noise, a multi-focus image fusion method based on HOSVD and edge intensity is proposed. And a novel edge intensity measure is proposed, which integrates the horizontal, vertical, 45°diagonal edge information and 135°diagonal edge information. Because a single image fusion strategy is easy to lose useful information, a multi-strategy fusion rules based on sigmoid function and edge intensity is proposed. And the shrink factor k is achieved by regional energy. The multi-strategy fusion rules can identify complementary and redundant information in images. Considering multi-strategy fusion has more advantages than a single fusion strategy, so the multi-strategy fusion rule based on sigmoid function is used in the following methods. The research indicates that for the multi-focus images, the two proposed methods achieve good results while the second proposed fusion method also can effectively suppress noise in images.(2) On the basis of Shift Invariant Shearlet Transform(SIST), two new image fusion methods are proposed. The first new method is proposed based on SIST. The different fusion rules are employed to fuse low and high frequency subband coefficients. The clarity indicators of structure tensor is used to identify low-frequency coefficients, a choose-max scheme based on the extracted feature is presented. To identify the details effectively, edge intensity is used to measure high-frequency subband coefficients. And a multi-strategy fusion rule based on sigmoid function and edge intensity is designed to fuse the high frequency subband coefficients. In order to improve the robustness of the fused image, an adaptive multi-strategy image fusion method based on riemannian metric is proposed. On the basis of riemannian space, a new dissimilarity is proposed. Affine invariant metric and Log-Euclidean metric are used to measure complementary and redundant information of the image. And an adaptive multi-strategy fusion rules is proposed and achieved by sigmoid function. Experimental results show that the former method achieves good result on multi-focus images and remote sensing images. While the latter method has wide application range, and performs good result on multi-focus images, remote sensing images, medical images and infrared and visible images.