Pixel-level Image Fusion Based On Non-tensor Product Wavelet

Image fusion integrates the information generated by different imaging systems about the same scene and objects or by the same sensor in different time according to some rules and creates a new image. The fused image contains a more complete, accurate and reliable description of the scene than any of the individual source image. As a result of this processing, the fused image is more useful for human and machine perception or further image processing such as object recognition, image analyzing, feature extraction and object trace. Image fusion has been widely used in automatic target recognition, machine vision, remote sensing, robotics, complex intelligent manufacturing systems, medical images and other non-military areas or military fields, and become a hotspot. In accordance with the level of abstraction of the image fusion process, the level of image fusion from low to high is divided into:pixel, feature, decision levels. Pixel level image fusion is to be processed directly on the pixel information to get a fusion image. Most of the fusion algorithms are concentrated at this level.The non-tensor product wavelet is the development of wavelet theory and its applications are lagging far behind the tensor product wavelet. As an important aspect of non-tensor product wavelet theory, the non-tensor product wavelet filter design has become a focus for researchers. Compared to the tensor product wavelet filter,the construction of non-tensor product wavelet filter has greater freedom and singularity characteristics to capture all directions, it also has all the capacity of traditional wavelet:multi-resolution analysis and low computational complexity. In this way, images decomposed by non-tensor product wavelet can provide more high frequency information, more conducive to the image information analysis, and greater application prospects. Linear phase (symmetric or antisymmetric) wavelet has perfect reconstruction characteristics, the type of wavelet image reconstruction does not produce phase distortion, and thus will not lead to the edge distortion, which is extremely important for image fusion, because the final step of image fusion process is often the image reconstruction, of all Daubechies family wavelet,only Haar wavelet has linear phase, therefore, constructed non-tensor product wavelet filter banks with linear phase,compactly supported, orthogonal is needed to image fusion.This paper studies the basic theory and methods of the construction of two-dimensional non-tensor product wavelet filter, gives specific examples of two-channel, three-channel, four-channel filter banks, and apply filter banks constructed in the previous chapter in image fusion experiments. On one hand, for multi-focus image fusion, the fusion method based on non-sampling symmetric non-tensor product wavelet is proposed. Information entropy, standard deviation, average gradient, and spatial frequency indicators are used to evaluate the fused images, the results show that the proposed method has better visual effect than the tensor product wavelet-based fusion method. On the other hand, for multi-spectral and panchromatic image fusion,this paper proposed a method based on IHS transform and the four-channel symmetric non-tensor product wavelet, the correlation coefficient, spectral distortion degree, the relative average spectral error are used an a performance evaluation of fusion images. The experimental results show that the edge details of the fused image of proposed method has better visual effects, and its ability to maintain the spectral information and high spatial resolution outperforms the improved IHS method, the DWT methods, and IHS-DWT method.Finally, the direction of further research is presented.