Image Registration Fusion Algorithm Based On CS And Mutual Information

Image is the most important source of information. With the development of informationtechnology, high quality of the image is needed, so that the amount of data acquired increaseslarger and lager, which brings great pressure to the data storage, transmission and processing.Recently, compressed sensing(CS) is proposed to sample and compress the signal as a whole,which can greatly reduce the amount of data and resolve this contradiction.Image registration is the basis and prerequisite for image fusion process, whichdetermines the effect of registration accuracy and quality of subsequent processing. While CSimage sampling techniques can relieve the pressure and the amount of data, it also brings newproblems o image registration: The outputs of CS is measurements of images, not imagesthemselves. Reconstruction is needed to restore the original image, before registration andfusion. In this paper, a new method, reconstruction-registration method is proposed, Thismethod considers the image reconstruction and registration as a whole, avoiding the repeatedtransformation between image and domain sparse domains, in order to improve the efficiencyof reconstruction and registration. The main work is as follows:（1）Based on the study of mutual information and image registration, investigation wasmade for image registration method based on wavelet transform and mutual information. Bydecomposition of wavelet transforming, image registration is conducted.（2）Under the CS frames, this paper is focused on the registration method of remotesensing images. What this method differs from the conventional method is that thereconstruction of CS, registration, and fusion are referred to as a whole, in which input is theoutput of the measurements of image sensor, rather than the traditional image. Waveletcoefficients is obtained by CS reconstruction, and then in the wavelet domain, registrationand fusion is conducted based on maximum mutual information registration criteria, so toavoid the repeated transforming between image domain and wavelet transform domains andimprove efficiency.（3）Simulation is made to prove the feasibility of the methods described above.Results show that our method can save hardware costs, and can use of prior knowledge toreconstruct the unknown signal quickly.