Application Research On Image Fusion Based On Statistical Modeling Method

As one important branch and research focus of data fusion, image fusion is anindispensable modern high-new technology in the field of image understanding and computervision, which is the synthesis and integration of sensors, image processing, signal processing,computer and artificial intelligence. The basic idea of the image fusion is: Through somecertain algorithms to combine two or more different images from multi-sensors into a newsingle image by makeing full use of the complementary and redundancy information of thesource images, which can provide more accurate, more comprehensive, more reliable andeasier to understand description about the same object target or scene than the source images.Besides, this process can increase the image’s authenticity and understandability whileimproving the accuracy and robustness of the observation.At present, as one commonly used kind of image fusion methods, the multi-resolutionimage fusion methods have always been a hot research topic and can be divided intomulti-resolution pyramid fusion methods algorithm and wavelet transform fusion methodsand the image fusion methods based on multiscale geometric analysis(MGA) transform.Among them, separable2-D wavelets only have a finite number of directions in the horizontal,vertical and diagonal, and cannot obtain the optimal or sparse representation about the “line”or “face” singularity of higher dimensional functions. To overcome these limitations, MGAtransform are proposed and they can perform better in describing image edge, contour, corner,directional texture and other important image features or details with high directionalsensitivity and anisotropy, which have broad application prospects in the field of image fusionand so forth. In addition, multiscale decomposition coefficients for natural images typicallydemonstrate strong dependencies across scales, within scales and different directions,therefore, it has very important theoretical significance and research values to studycomprehensive statistical models to capture dependencies between coefficients moreaccurately and efficiently in the multiscale transform domain.On the basis of learning typical image fusion algorithms and existing statistical models,we have made a further research on the establishment of multiscale statistical modelingmethods in the paper. Given these study results, its applications in image fusion are exploredand studied more in depth. The main content of this paper is organized as follows:(1) Through the study about multi-scale geometric analysis theoretical knowledge and itsapplications, we deeply analyze the strong dependencies across all of scales, intra-scale andinter-direction between the decomposition coefficients. Given that each coefficient is closelyrelated with its Generalized Neighborhood Coefficients (GNC), a new design procedure togenerate context by combining GNC coefficients has been developed. Then, a comprehensivestatistical correlative model is constructed by integrating context information with HMM,which can fully and accurately describe the correlations of decomposition coefficients.(2) Based on the new context design solution, a novel contextual hidden Markov model(CHMM) and modified Pulse Coupled Neural Network (M-PCNN) based fusion approach inthe Contourlet domain is proposed for multi-modal medical image fusion. The high-frequency directional sub-bands are highly non-Gaussian distribution (mean a sharp peak at zeroamplitude and heavy tails to both sides of the peak) and exhibit clustering and persistence.Considering the powerful advantages for statistical modeling and processing of HMM, thenew context design scheme integrated with HMM is established to construct a comprehensivestatistical correlative model, which can collectively capture persistence across scales,directional selectivity within scales and energy concentration in the spatial neighborhood ofthe coefficients. Then calculate the modified spatial frequency (MSF) as the feeding inputs tomotivate M-PCNN and utilize CHMM model parameters to obtain the values of link strength.Next, count the accumulative total of firing times and select the high-frequency fusedcoefficients by maximum strategy while low-frequency sub-band coefficients are fused by themagnitude maximum rule. We apply different fusion rules to the Contourlet coefficients athigh and low frequency sub-bands. The simulation experimental results demonstrate that,compared to other fusion methods, this novel algorithm can effectively integrate usefulinformation of the original images, enhance image contrast, preserve image contours, richimage details, and greatly improve image sharpness and quality.(3) With the new context design procedure, a multi-focus noisy image fusion methodbased on sharp frequency localized Contourlet domain contextual hidden Markov model(SFLCT-CHMM) and improved dual-channel pulse coupled neural networks (IDPCNN) ispresented. In the digitization and transmission process, images maybe affected by imagingdevices, light, temperature and other external factors, this led to the introduction of noise.The subsequent procedures of image fusion research will be affected if they are not eliminated.Therefore, choosing an efficient and reliable image denoising algorithm is a pretty importantpreprocessing step. Firstly, SFLCT-CHMM statistical model is established and applied toimage denoising. Multi-focus noisy images are decomposed into sub-images by SFLCT,solving one drawback of the original Contourlet transform that its basis images are notlocalized in the frequency domain. Meanwhile, we introduce cycle spinning (CS) technologyto compensate for the lack of translation invariance property of SFLCT and suppress thepseudo-Gibbs phenomena around singularities of images. Experimental results show that thedenoised image can be gained with preserving image edges, contours and other details whileeffectively removing noise for higher PSNR values. Next, the fused image can be obtained byusing a novel fusion method based on IDPCNN. IDPCNN has powerful biologicalbackground, thus it will better protect image texture information and improve the resolutionand accuracy of the fused image. Through the analysis on the experimental results ofmulti-focus noisy images, the presented fusion method can effectively reduce the effect ofnoise, and further improve objective evaluation and visual effects, which verify its feasibilityand effectiveness.