Application On Image Fusion Of Weighted Gradient-based Algorithm With Steering Kernel And Structure Tensor

In recent years, significant attention has focused on multi-source image fusion for bothmilitary and nonmilitary applications. Multi-source image fusion techniques combinerelated image information from multiple sources of sensors to achieve more informativefused image than the one could be achieved by using a single, independent sensor.Currently, the types of image sensor are numerous. The system of image fusion theories arefar from becoming mature and uniform. Designing reliable and real-time fusion algorithmsby taking account of the inherent characteristics of each type of imageries is one of thecurrent research focuses.This thesis focuses on the study of fusion algorithm for multi-spectral images and anovel weighted gradient-based fusion method combining structure tensor with steeringkernel is proposed, which prevents blur, artifacts and mis-registration and sharpens edgessuccessfully. In addition, lots of complicated mathematical calculations are avoided. Themajor results of the research are as follows:1. Research on the multi-spectral image gray fusion algorithm through variationaltechniques to achieve a more formative gray image. We combine the gradients ofmulti-spectral images to obtain the geometry of the fused image. In the variational fusionframework, the term of the weight for source images is imported to enhance the perceptualcontract of the fused image.2. The new structure tensor based steering kernel regression for irregular interpolationalgorithm is applied in image fusion. First, the latest kernel regression algorithms arecovered and the defects and limitations of self-adapting Bilateral Kernel Regression andSteering Kernel Regression are analyzed. Then a novel weighted gradient-based fusionmethod combining structure tensor with steering kernel for multi-spectral images isproposed. The structure tensor is usually used to extract the main gradient formation of asingle point from the source image. Therefore we propose to use steering kernel to describethe spatial structure information of the source images, which is robust to the noise. Finally,the structure tensor is combined with the steering kernel though a saliency metric.Experimental results demonstrate that the proposed method outperforms other conventional fusion methods in term of visual comparison and quantitative assessment.