Research On The Registration And Fusion Of Multi-sensor Sequence Images

Nowadays, the single sensor image is difficult to meet the current requirements. Meanwhile, with the rapid development of multi-sensor, the quantity and the diversity of images are growing. Multi-Sensor sequence images are from different sensors, and every sensor creates a series of images. As a result, multi-sensor sequence images’ processing plays an important role in the field of theoretical study and practical application, such as digital image processing, target recognition and so on.The whole study in this paper involves four parts:multi-sensor images registration and fusion, sequence images registration and super-resolution reconstruction. The specific work is presented as follows:(1) Based on the basic theory and methods of multi-sensor images registration, according to the differences in gray and resolution, a registration algorithm is proposed, which is based on the contour information. Firstly, well-defined main contours are obtained and polygon fitting is performed after detecting. Secondly, the contours are indicated by the Freeman chain code and aligned. The polygon vertices along the matched contours are used as control points. Experiments show that the algorithm can not only solve the multi-sensor images registration problem fast and accurately. Besides, this algorithm can be applied in single sensor images’field.(2) Based on the basic theory and methods used commonly of multi-sensor images fusion. A multi-scale image fusion scheme based on the wavelet transform is presented in this paper. The wavelet transforms of the source images are computed, which contain four bands at every scale. Then different integration rules are selected to fuse them correspondingly. At last, the fusion result is obtained by inverse wavelet transform. The contrast experiments show this fusion algorithm’s advantage.(3) Based on the methods of image registration and the theory of super-resolution reconstruction (SRR), an algorithm of sequence images registration is proposed in this paper, which combines the feature with the gray information of the images. The whole process is divided into two steps:firstly, align roughly based on the contour, secondly, align accurately based on the gray. Choosing the normalized mutual information to construct the objective function and optimizing by the Powell algorithm are the innovations of the second step. The rough parameters are chosen as the iterative initial value to avoid optimization results over dependence on initial estimations. Experiments show this algorithm get accurate parameters and work well even though the images move greatly.(4) By analyzing the image degradation model, the ill-posedness of the SRR and some spatial domain methods, a regularized SRR algorithm based on MAP criterion is proposed to achieve multi-frames SRR. Based on the analysis of the Bayes theorem, the cost function is deduced by which that Gauss random field is used to estimate the conditional probability and Gauss-Markov random field is used to estimate the prior probability. Then the reconstruction algorithm is built by designing regularization item. Experiments show the algorithm is easy to realize and perform well in preserving the edges of images.For the multi-sensor sequence images, the image registration includes the multi-sensor images registration and the sequence images registration and the image fusion includes the multi-sensor images fusion and the SRR in this paper. The algorithms this paper proposed solve the registration and fusion problem effectively, and provide a new technical approach for the fast, robust, universal and accurate registration method and efficient and robust fusion method by which the image registration and fusion could be developed.