Pixel-level Fusion Algorithms On Multi-focus Image Fusion

The thesis is devoted to studying pixel-level multi-focus image fusion algorithms. The important existing schemes for multi-focus image fusion are summarized and analyzed. Some solutions to the problems caused by the existing schemes are introduced in space domain and transform domain. Some kinds of image fusion experiments are performed and the fusion results are evaluated quantitatively and objectively in comparison with other schemes, so some valuable conclusions are obtained. Moreover, an overall introduction about the elementary knowledges of image fusion, the characteristics of multi-focus images and the relevant theoretics is done in this thesis.Firstly, this thesis explains the background and meaning of the subject, and summarizes the development and main application areas of image fusion.Then this thesis depicts imaging mechanism of multi-focus images, and the survey of multi-focus image fusion from several aspects, which refer to basic theories, fusion schemes and image fusion performance measures. Multi-focus image fusion is usually performed at one of the three different processing levels: pixel, feature and decision. Pixel-level multi-focus image fusion is the subject matter of this thesis. The goal of pixel-level image fusion is to represent the visual information present in input images without the introduction of distortion or loss of information. In practice, the goal is modified to: the fusion, or preservation in the output fused image, of the"most important"visual information that exists in the input image set. The relevant fusion schemes can be categorized into space domain and transform domain. The fusion algorithms in space domain, including logic filtering algorithms, arithmetic algorithms and partition fusion algorithms,produces the fused image pixel by pixel directly, without any transform to registered images. The fusion algorithms in transform domain decompose input images at various resolutions, fuse individual or groups of pixels from the multiresolution pyramid representations with different schemes. This thesis studies multi-focus image fusion algorithms in space domain and transform domain. Then multi-focus image fusion performance measures are introduced, the performance should be measured with subjective and objective methods, i.e. the objective evaluation of fused image quality should be complemented based on subjective evaluation.Image fusion algorithms in space domain are studied. Because arithmetic fusion algorithms suffer from low contrast and low correlation between pixels, and partition fusion algorithms suffer from improper partition and incorrect choice of partitioned sub-blocks, two kinds of improved nonuniform partition fusion algorithms are introduced. Calculating characteristic of sub-blocks using formula (3.1). Nonuniform Partition Algorithm One chooses or further partitions sub-blocks based on the comparison between gradients of sub-blocks to be fused, calculated by formula (3.5). Nonuniform Partition Algorithm Two partitions the images to be fused, which are blurred with Gaussian function, then chooses or further partitions sub-blocks based on changing gradients of the initial and blurred sub-blocks, which is explained in section 3.4.2. The experimental results show that compared with uniform partition algorithm and"the nonuniform partition algorithm"proposed in 62 refs, the fusion image quality using nonuniform partition algorithms is slightly better, especially in"complex areas", which is explained by figure 3.3 (k), (l), (m), (n). As far as the two nonuniform partition algorithms, the computation complexity of Algorithm One is lower, the fusion quality of Algorithm Two is better. Nevertheless, the two improved nonuniform partition algorithms can only lead toreduction of blocking effect, without avoiding it, which is the primary problem of partition fusion algorithms. It can be concluded that although partition algorithms helps to preserving the important information from input images and increasing the correlation between pixels, it suffers from being scant of detail expression.Image fusion algorithms in transform domain are studied. The human visual system is primarily sensitive to local contrast changes, e.g. the edges or corners, which makes fusion algorithms with multiresolution pyramid analysis (e.g. Laplace pyramid, ratio pyramid etc) replace algorithms based on space domain gradually. Data decomposed by pyramid analysis is redundant and inherently correlated, and scant of orientation selectivity. Discrete Wavelet Transform (DWT) provides a good time-frequency analysis of signal, and results in a non-redundant signal representation and an optimal representation for singularities, so DWT is widely used as a tool of multiresolution analysis. Nevertheless, DWT suffers from four fundamental, intertwined shortcomings: oscillations, aliasing, shift variance, and lack of directionality. The Dual-Tree Complex Wavelet Transform (DT-CWT), which is a simple solution to these four DWT shortcomings, can represent detail features of images. Therefore, as a tool of multiresolution analysis, DT-CWT is adopted in this thesis. The existing widely used window-based algorithms are summarized from formula 4.23 to formula 4.28. The fusion rule of low frequency coefficients should represent the image definition, and ensure that the chosen pixels come from the clear areas as many as possible, so the ones having maximal neighborhood gradient are selected. Experience results show that this method well preserves image contour information. High frequency coefficients represent edge and detail information. The values of high frequency coefficients in clear areas are bigger than that in blurred areas, so the ones having maximal absolute values are selected. Experience results show that the proposed algorithm, which is called"AMulti-Focus Image Fusion Algorithm with DT-CWT", not only solves the problems such as low contrast and blocking effects caused by fusion algorithms in space domain, but also avoids the artifacts and ringing artifacts exhibited by conventional wavelet based fusion algorithms. Furthermore, the coefficient fusion rule proposed in this thesis also suits the fusion of DWT coefficients, and improves the fusion image quality. Review the thesis, the creative and valuable work as following:1. Improved nonuniform partition fusion algorithms are proposed.2. A multi-focus image fusion algorithm with DT-CWT is proposed. For the low frequency coefficients, the ones having maximal neighborhood gradient are selected. For the high frequency coefficients, the ones having maximal absolute values are selected.