Research On The Applications Of Edge-preserving Image Filtering Methods

With the development of computer,communication and sensor technologies,the rapid acquisition of massive and high-quality image data becomes possible in various fields such as remote sensing and biomedicine.A breathtaking digital image world has been formed,and a new era of digital image application is approaching.Almost accompanied by the appearance of digital images,the image filtering technology is invented.Image filtering is a neighborhood operation,in which the output value of any given pixel is determined by applying some algorithms to the values of the pixels in the neighborhood of the corresponding input pixel.Image filtering is one of the most basic operations in image processing,and also the basic step involved in most computer image information extraction.After nearly 30 years of development,the edge-preserving image filtering technique,has been widely used in image denoising,edge detection and some other classical problems in the digital image processing field.Meanwhile,with the introduction of various new theories,and the theories and methods of the edge-preserving image filtering have also been further developed,giving rise to new application fields,playing an increasingly important role in computer vision and other fields.The appearance of massive image data and the demand of a great deal of new image processing and analysis have also raised new problems and challenges for the image filtering.This research reviewed the existing edge-preserving image filtering methods and chose the mean shift filtering and guide filtering as theory tool.The tool is employed to solve the typical problems in the natural scene images or remote sensing image processing.The research works were carried out from multiple image processing levels,which involved in remote sensing image fusion,image segmentation and remote sensing image feature extraction and classification.The whole research work can be summarized as the following:1.The research systematically summarizes various edge preserving filtering methods and applications,and analyzes the research trend.The two typical edge preserving filtering methods with superior performance,i.e.,the mean shift filtering and guided filtering,are discussed preliminarily from the aspects of theoretical basis,parameter setting and method application.2,The guided filtering is introduced into remote sensing image fusion process to effectively alleviate the spectral distortion problem.Image fusion,also called as image pansharpen,is the process of combining the image data of the same scene acquired by multiples sensors into a single high-quality images with the advantages of each source.The fusion of remote sensing a panchromatic image and a multispectral images,also called as image pansharpen,is expected to generate a new image with both rich spatial details of the panchromatic image and the spectral information of the multispectral image.The component substitution based methods(CS)represent a group of classic pansharpen methods,which consists of two crucial steps,including the simulation of the low-resolution panchromatic band and the detail injection.These methods feature low computational complexity and the improvement of the spatial resolution of images after fusion,but the fused results also suffers from serious spectral distortion.The main cause of the spectral distortion is the mismatch of luminance distribution between the replaced and replaced bands during the process of panchromatic band simulation.In order to solve this problem,a local adaptive component substitution fusion method is proposed,in which the local adaptive model is used to replace the original global simulation and detail injection models.The first block is to simulate a low-resolution panchromatic band by a local linear regression model between panchromatic and multispectral bands.The second block extracts spatial details and adds details back to multispectral bands in locally varying ratios.By recasting the local linear regression model into the guided filtering framework and analyzing the implicit statistical assumptions underlying CS methods,the strengths of local based pansharpening algorithm are addressed.The method makes full use of the structure-transferring property of the guided filtering,can effectively transfer the structure of the guide image to the input image,and thus,simulate a proper low-resolution intensity band.Experiments test 7 pairs of images acquired from different sensors,such as GF-2,QuickBird and Worldview-2.Both quantitative and qualitative evaluations reveal that the presented method can better preserve the spectral information than some states-of-the-arts methods.3.During the process chain of image understanding,regional segmentation is a bridge between the pixel space and the feature space,and the capability of the segmentation method directly affects the efficiency and accuracy of subsequent image interpretation.The mean shift segmentation algorithm is an extension of mean shift filtering algorithm and has been widely concerned in both natural and remote sensing image analysis.However,this method do not make use of the discontinuous information over local object boundaries,and it also fails to effectively detect and segment objects with weak edges.To solve this problem,a mean shift segmentation framework imbedded with a boundary confidence measurement is proposed,and the boundary confidence map is adaptively integrated into the iterative process of the mean shift segmentation in the form of weight.In order to obtain reliable boundary confidence map,multiple edge detectors are integrated by regression models.The first group of detectors consists of gray gradient and the angle measurement between the gradient vector and the ideal edge vector,the second group of detectors consists of brightness gradient,color gradient and texture gradient.The two groups of detectors are connected to Logistic regression and support vector regression(SVR)models respectively,and results in 4 different boundary confidence estimation algorithms.The proposed framework and methods are trained and predicted through the Berkeley Benchmark300 dataset.The segmentation quality is evaluated by 4 quantitative indicators:probability random index(PRI),global consistency error(GCE),variation of information(Vol)and the empirical evaluation function for color image segmentation evaluation(EEF),as well as visual interpretation.Experimental results show that the integration of the boundary confidence can alleviate the under-segmentation problem of the general mean shift segmentation.The combination of brightness gradient,color gradient and texture gradient with SVR can obtain the overall optimal boundary confidence prediction performance,and then obtain the best segmentation performance.With the integration of the boundary confidence information,the boundary of the texture objects or the objects with weak gradient can be detected accurately.4.Taking the over-segmented regions obtained by the mean shift segmentation as the basic analysis units,this research proposes a large-scale classification remote sensing scene by combining regional multiscale segmentation and Markov random field(MRF).Multiscale analysis has been widely used in feature extraction and the modeling of high-resolution remote sensing images.The subsampled wavelet transforms are commonly employed for establishing the multiscale representation of an image.However,the wavelet-based features cannot describe patterns with long spatial span,and often result in noisy classification results.In contrast,the object-based image analysis(OBIA)can create classification maps composed of compact land objects.But features extracted from a single scale still cannot provide discriminating information for the land cover classification.To improve the classification accuracy,meanwhile,alleviate noisy thematic maps,a regional multiscale classification method is proposed.In the first block,the mean shift segmentation method is employed to create initial over-segmented regions.Thereafter,a rule combining the gray values in the regions and the shared boundary lengths among regions is designed to extract the low-frequency part of the image.The current high-frequency part is obtained by the subtraction of the original image and the current low-frequency part.By replacing the original image with the low-frequency part,and repeating the segmenting and decomposing process,a regional multiscale representation can be iteratively established.In the second block,the classification result obtained from the original image is seemed as the prior of the label field in the first decomposed level,and the high-frequency part in the first level models the feature field of MRF.The classification result in the level is obtained by solving an objective function consisting of the feature and label energies.Through iteratively projecting the current classification result to the next level and modelling the feature field with the high-frequency part,the final classification map can be obtained in the coarsest scale.The experiment conducted on synthetic texture images and multispectral remote sensing images shows that the mean shift segmentation algorithm integrating boundary confidence information can effectively provide analysis units for the following region modeling process.Compared with the model established on the multiscale wavelet domain and the single-scale regional model,the proposed method can effectively improve classification accuracy and avoid the generation of "pepper salt" phenomenon.5.During the land cover classification based on the hyperspectral remote sensing images,the feature extracted from single scale cannot effectively express the difference between land classes and distinguish the boundary of ground objects.In order to solve this problem,based on the observations that multiscale abstraction of land surface can be achieved by applying a series of scale parameters on the edge-preserving filters,both multiscale mean shift and guided filtering features are extracted,and are expected to benefit the following image classification process.Firstly,principal component analysis is used to reduce the dimension of hyperspectral images.Then,the first principal component or the first three principal components are taken as the guiding image,and the first several principal components with the most information are taken as the input image respectively.The multi-scale features can be extracted by guiding filtering with multiple scale parameter settings to represent the structural information of land objects at different scales.Meanwhile,the spatial domain and range domain bandwidths with ascending order are also set to extract mean shift based multiscale features.Finally,the features extracted from different scales are stacked to form feature vectors and input into the classifier for supervised classification.Three hyperspectral datasets,including the Pavia University(University)and the Centre of Pavia(Centre)and Salinas,were used to extract the multi-scale guided filtering features,multi-scale mean drift features,extended morphological features(EMPs)and multi-scale Gabor texture features,and input them into the support vector machine(SVM)and random forest classifiers.All the features and performance of the classifier are compared and analyzed.The experimental results show that the multiscale feature fusion achieved feature stacking is helpful to the improvement of the overall classification accuracy,but usually at the cost of the reduction of the classification accuracy of individual land classes.Compared with EMPs,the two proposed multiscale features can effectively retain the boundary structure of ground objects while smoothing the details of ground objects,and obtain better classification results on multiple scales.The experiment verifies the validity of the proposed multi-scale features.