| With the rapid development of imaging device,it is very convenient to capture high-resolution images.However,due to the inherently limited depth-of-field in the image sensor,the objects in the scene only at a particular distance from the camera are in focus while the others are defocused.The defocus effect of image is manifested by varying degrees of blurs and loss of detailed information.In some practical cases in engineering application and biomedical research,when the objects of interest or specimens span a large depth range,low depth-of-field is not desired.Multi-focus image fusion(MFIF)is an inexpensive yet effective technique to tackle the issue of limited depth-of-field of image sensor.Several images of the same scene captured by different focus settings are fused using some algorithms to generate a composite image,in which all parts are fully in focus.The composite image should contain more detailed information than any individual source image,which will improve the interpretability of the data.Thus,it is more suitable for human visual perception and subsequent image processing tasks.At present,MFIF technology plays an important role in many image processing-based applications,such as long range surveillance,light-field camera,microscopy imaging,etc.The two most challenging and necessary tasks of MFIF method are to accurately evaluate the focus properties of pixels and to identify the best focused regions of each source image.This dissertation concentrates on the study of these two key issues and four novel methods based on the characteristics of multi-focus images are proposed.The specific research contents are:1.The traditional multiscale decomposition(MSD)based image fusion method cannot effectively preserve detail information and easily decrease contrast of the fused image.Faced with these problems,we present a novel MFIF method based on multiscale shearing non-local guided averaging filter(MSNLGA).First,we construct a new multiscale geometrical analysis(MGA)tool called MSNLGA,which combines the non-local guided averaging filter with the shearing filter bank.The MSNLGA can represent the intrinsic geometric structure of image sparsely due to the combination of the feature of preserving edge information by the non-local guided averaging filter and the ability of capturing directionality of images by the shearing filter bank.Then,the MSNLGA is used to decompose source images to obtain approximate subbands and directional detail subbands.Since the approximate subband is different from the directional detail subband in terms of characteristic information.Two different fusion rules are designed for the approximate subbands and directional detail subbands,respectively.For the approximate subbands,we calculate the spatial frequency(SF)feature by using the discrete Tchebichef orthogonal polynomial transform(DTT)coefficients.The SF feature based on DTT domain has a strong anti-noise ability.For the directional detail subbands,we introduce the convolutional sparse representation(CSR),which is a model that can achieve sparse representation of an entire subband and has shift-invariant property,to represent each subband so as to reduce the influence of mis-registration on the selection of coefficients.The experimental results show that the fused images of our method can well retain the detailed information,and can still accurately select the fused coefficients under imperfect conditions(such as the noise,mis-registration).2.To address the weaknesses of spatial inconsistency and inaccurate boundary positions of the fusion decision maps generated by the existing fusion methods based on spatial domain,a robust MFIF method based on lazy random walks(LRW)and multiscale focus measures is presented.Inspired by the fact that human perceives the blurriness of an image patch is closely related to the scales of patch,a multiscale sum of the weighted modified Laplacian(MSWML)operator is firstly adapted to extract the focus information of source images at different scales.Then,the scale-consistent focus maps are computed by collaboratively employing the local focus information as well as the global focus information.The information of them indicates the stable and accurate focused regions of source images,but they are sparse and cannot use in the fusion step directly.After that,the scale-consistent focus maps are fed into LRW-based estimation framework to produce the dense weight maps.Finally,a weighted-sum strategy is used to generate the fused image.Experimental results show that our method can precisely locate the boundaries of focused regions and has state-of-the-art performance for MFIF under various situations encountered in practice.3.Inspired by the important guiding significance of color information in perceiving image quality,we explore the use of two unique characteristics of color images(color structure information and depth information)for color MFIF tasks,and a color MFIF method using a quaternion morphological gradient and improved KNN matting is presented.We first propose a new focus measure called the quaternion morphological gradient,which is derived based on the quaternion representation of color images and a proper ranking function,for evaluating the sharpness of the source image.The phase information of the Clifford translation of quaternion is used for determining the order of two quaternions.Then,the focus information of source images is used to produce the initial decision maps.In the decision map refinement step,we add the depth information of source image into the feature vector to hopefully improve the robustness of KNN matting results.Experiment results show that the fused images of our method can maintain the detailed information in the source images very well.Even if the source images have obvious mis-registration,the proposed method can still produce high-quality fusion images.4.Most sparse representation(SR)-based fusion methods use the manner of separately handling color channels,which easily cause hue distortion and color saturation reduction in fused images.Quaternion sparse representation(QSR)employs the quaternion matrix to model color image in a holistic way,and it can fully exploit high correlations among color channels and well preserve the inherent color structures of source images in reconstruction result.We propose anovel color MFIF method based on QSR to address the issues of traditional SR-based fusion methods.In this method,we first learn a clear quaternion dictionary on a high-quality image set and a blurry quaternion dictionary on a Gaussian blurred version of the same set by employing the K-quaternion singular value decomposition method.Then,the clear quaternion sparse coefficients and the blurry ones of each image patch are estimated by a simultaneous quaternion orthogonal matching pursuit algorithm and quaternion orthogonal matching pursuit algorithm,respectively.After that,the salience feature and sparsity feature of each image patch are computed from the clear quaternion sparse coefficients and blurry quaternion sparse coefficients,respectively.Now,we design a new activity measurement of each patch by exploiting both salience feature and sparsity feature.Additionally,the computation process of the two features fully considers the spatial information to enhance its discriminant power.Finally,a maximum fusion rule is adopted to combine the clear quaternion sparse coefficients of source images,and then we can precisely reconstruct the fused image using the fused sparse coefficients and the clear quaternion dictionary.The experimental results demonstrate that our method successfully avoids the color distortion in the fused images and performs favorably against some recent SR-based fusion methods qualitatively and quantitatively. |
