Variational Partial Differential Equations-based Image Fusion Method

In recent years, the variational partial differential equation (PDE) based image processing methods are widely used and have made great process in the field of image processing. The purpose of image fusion is to combine useful information from different source images of the same scene to achieve a new image which is more correct clear and complete knowing about the target than a single source image. Research shows that the distinct characters of an image could be represented by its geometric informations. Using variational functional models can easily deal with the gradient, curvature, and other important information of the image, and can preferably preserve the local feature of digital image, making the processed image has a perfect vision effect. So the variational PDE algorithm is used into image fusion is feasible.The main research work and innovations of this thesis includes the following aspects:(1) the pulse coupled neural networks (PCNN) is a novel artificial neural network, which is presented for emulating the visual neurons activity, PCNN has been successfully applied in image fusion. In this thesis, PCNN models and its actual application of image fusion are studied, and PCNN is combined with variational PDE.(2) At certain extent, grayscale image could be depicted as the gradient module, the greater the gradient module, the higher the image sharpness and the richer the image detail. According to this characteristic. This thesis presents a new variational model based on the square of gradient module, which making the largest square of gradient module of fused image. In order to keep there is not much difference between the fused image and source images, Fidelity term is used in the variational model, and use PCNN to calculate the contribution of the source images in fidelity term.(3) Similar to grayscale image, at certain extent, vector image could also be depicted as the gradient module of the vector image. This thesis describes the source images as component images, and describes the fused image as a vector image:to all vector images which could made up of several component images, the fused image has the largest gradient module, a new variational model based on the gradient module of vector image is presents. Fidelity term is also used, and here also use PCNN to calculate the initialization values of the fused image in fidelity term.(4) To image fusion method based on multiresolution analysis, we set focus on the wavelet transform and contourlet transform, and some common fusion rules. In terms of fusion rules of low-frequency domain, to solve the shortage of the direct average method, this thesis presents a new fusion rule for low-frequency domain, which is based on the gradient module of vector image.To test the validity of the algorithms in this thesis, many simulation examples are demonstrated. Test results show that our algorithms are effective and feasible in both objective quality evaluation and the visual effect.