Region-based Multiresolution Image Fusion Algorithms

This thesis focuses on region-based multiresolution image fusion algorithm which is an intermediate fusion level from pixel-level towards aspect-level fusion. Region-based fusion approaches not only retain the accuracy of pixel-level fusion, but also they may help to circumvent some of the well-known drawbacks of pixel-based techniques, such as blurring effects, high sensitivity to noise and misregistration. The paper presents several novel region-based image fusion algorithms and obtains better fusion performance than pixel-level fusion. The main work can be summarized as follows:1. Chapter 1 depicts the background, objective and meaning of region-based image fusion. Besides, the chapter also introduces the development and actuality of region-based image fusion and presents the scheme and expectant results.2. This thesis is concerned with region-based multiresolution image fusion methods. So, the paper introduces the general framework for region-based multiresolution image fusion presented by G Piella[17] and explains each module of the framework in detail. Some common methods and rules of modules are given as reference. Moreover, this paper addresses some existing objective fusion performance measures because the evaluation of fusion result is an important component of image fusion.3. This paper denotes the disadvantages of the algorithm in literature [19], on this foundation, a novel region-based multiresolution image fusion approach with wavelet transform is presented. First, the multiresolution architectures of the input images are obtained using the wavelet transform. The low-frequency approximation images of input images are segmented based local region contrast and so we obtain a segmented image consist of legible region, blurry region and boundary between both regions. Then the high-frequency detail images of different region are fused by different fusion rules. And the composite approximation image is achieved by averaging the approximation of the input image. Finally, the fused image is reconstructed using the inverse transform. The experimental results show that the presented fusion scheme can achieve better fusion performance than pixel-level fusion method and the method in literature [19].4. The Discrete Wavelet Transform has two main disadvantages: lack of shift invariance and poor direction selectivity for diagonal features. The Dual-Tree Complex Wavelet Transform owns many excellent properties: approximate shift invariance, good directional selectivity, perfect reconstruction, limited redundancy and efficient order-N computation. So this thesis introduces the construction and properties of the dual-tree complex wavelet transform. On this foundation, a novel region-based multiresolution image fusion algorithm with dual-tree complex wavelet transform is proposed. The algorithm employs simple block segmentation method to segment the high-frequency detail image of input images. To the detail coefficients, the algorithm employs a block area variance as activity measure, and then the composite detail images are constructed by a selective combination where the composite coefficients belonging to a certain block come from the same input detail image, namely, the one whose activity in that block is higher. And the composite approximation image is constructed by the average of the input approximation images. First, it gives prominence to feature of object in image when region-based image fusion method remains the accuracy of pixel-level fusion. Secondly, due to shift invariance property and good direction selectivity, the dual-tree complex wavelet transform can preserve edge information and texture information. So the method proposed in this paper can obtain good performance, at the same time, improving the efficiency of DT-CWT.5. Chapter 5 introduces a new multiscale analysis named Nonsubsampled Contourlet Transform (NSCT). The NSCT is a complete shift-invariant directional multiresolution image representation. The NSCT is flexible in that it allows any number of 2l directions in each scale. Based on NSCT, a new method for region-based image fusion is proposed. First, the algorithm use NSCT to represent the input images at different scales and employs multifeature Fuzzy C-Mean (FCM) clustering algorithm to partition the approximation images at these scales. In this way, we obtain a multiresolution segmentation which is used to guide the subsequent fusion process. Then each directional detail image at each scale is fused by a rule that select the region which has higher activity (area contrast) as the composite image corresponding region. Finally, the paper compares the result of method presented in this thesis with others. Although having increasing computation, the experimental results indicate that images achieved by the proposed method have good qualities, especially for the IR image and TV image.6. In the end, the research work in this thesis is summarized, and the problems and disadvantage of current research work are also illustrated. Moreover, the paper discusses the recommendations for further work as well as the prospect of image fusion.In summery, image fusion is the focus concerned not only in academic study but also in practical application. Image fusion is an integrated technology which encompasses many theories, techniques and tools. Although the research presented in the thesis is limited, the author has done some beneficial attempts in region-based image fusion field. The deeper study in image fusion is necessary for us in future work and it must be a long-term process with more efforts.