Segmentation Of Highly Corrupted Near Bank Images By Impulsive Noise

Intelligent devices with vision systems working in or above surface of waters, haveapplication in military affairs, environmental conservation and scientific research. Whenthese devices work, the recognition of near bank images must be met. The water surfaceenvironment is not a simple configuration condition but no rigid environment. Varietyin direction, angle, intensity of illuminant and generation of a mass of random structureby trees and water will generate many changes in image detail such as a distortion ofedges and variation of pixels. So the image segmentation and recognition is verydifficult, it is much more difficult that water bank images segmentation and recognitionis. Two rough problems, more efficient filter to suppression impulsive noise from highlycorrupted images and segmentation of scene from water surface and sky of near bankimage, need be solved. In the process of research, three creative progressions areobtained.Discrepancy Measure between pixels is proposed, and two filters are proposed as well based on DM;Progressively reduced ordering vector median filter is proposed, and discrepancy measure progressive reduced order half switch vector median filter is proposed as well to filter corrupted images by random noise, especially in filtering highly degraded images by impulsive noise. It is well suit to preprocess;Drops of Water Region Growing algorithm is proposed to segment bank part from sky.The research process is shown as follows:1. Median and mEan Compromise Filter—MECFFirst, the part of noise and outlier samples is removed from considering order basedon median to obtain maintain order. Second, the filter output is an average value of allsamples in the maintain order. In order to evaluate effectiveness of MECF, thecomparison of the new filtering method with the relevant filters, AEM (Arithmeticmean method filter) and SMF (Scalar Median Filter), is employed. PSNR, SNR, MAE,MSE, NCD, NMAE and NMSE have used as quantitative measure for evaluationpurposes. The results show that MECF outperforms the other filters in terms of usedobjective criteria. But when corrupted pixels rate is more than 40%, the new filterbegins to show disadvantage.2. Discrepancy Measure between pixels—DMDM between a pixel and its neighborhood pixels is defined bywhere G(k (?) denotes the color pixel by a vector, ? denotes a vector λ-norm,k is the )dimensionality color images。It is a comparable metric to express deference degreebetween the pixel and other pixels within its neighborhood, so it is known if the pixel isa corrupted one. Then two main parameters of DM are optimized to different images,and two new filters are proposed based on DM. 3. Discrepancy Measure to Noisy pixel Vector Median filter—DMNVM First, it is estimated based on DM, whether the center pixel has been a corruptedpixel. If the central pixel is corrupted, it will be processed by VMF; and if the centralsample is noise-free, it remains unchanged. This requirement is satisfied by thealternation between the VMF and the identity operation rule based on DM. Theperformance of DMNVM is compared with well-known vector standards such as theVMF, vector median filter of minimum distance between central sample and scalarmedian MMVM, SMF, minimum distance between central sample and scalar median inits neighborhood vector mean filter FEVM and the basic vector directional filter VDF.The filtered images show that the VMF and MMVM excellently suppressed impulses inthe image, however some edges and image details are blurred especially on transitionsbetween objects; this undesired filter behavior is more visible in SMF, FEVM and VDF;the output of the proposed method shows the excellent signal-detail preservation withthe simultaneous impulse noise suppression. Moreover the results show excellentrobustness of the proposed method in terms of used objective criteria, especially inNCD, the minimum value of other methods is 18.6 times as much as that of theproposed method. Finally, the values of objective criteria—MAE, MSE, NCD ofDMNVM are 1.81, 2.59, 2.17 times less than that of the adaptive vector median filter. 4. Discrepancy Measure Switch Median filter—DMSM If the corrupted image pixels rate is low, the effectiveness of DMSM is as same asthat of DMNVM; if the image is in high noise level (>40%), the results show that theproposed method outperforms others (VMF, PSM—Progressive Switch Median filter)mentioned in terms of used objective criteria—PSNR, SNR, MAE, MSE, NCD,NMAE, NMSE, for example, when the corrupted pixel rate is 90%, its PSNR enhances6.71 more than other methods, and its NCD is 4.32 times less than the minimum ofother methods. Comparing with same type method--self-adaptive algorithm ofimpulsive noise reduction in color images (SAINR), when corrupted rate is less than40%, the PSNR and SNR of SAINR have a little advantage but MAE reverses. If thecorrupted rate is more than 40%, three objective criteria—PSNR, SNR, and NCD ofDMSM have the advantage over that of SAINR. The computational time of DMSM is0.63 times as much as that of VMF averagely. 5. Progressively Reduced ordering Vector Median filter—PRVM If all pixels of moving window are considered as a set, its subset that minimizes thevariation range of gray values in per channel must be gotten. The output of PRVM is thevector median of the subset. In terms of PSNR, if the corrupted rate is more than 40%,PRVM has the advantage over VMF, and if the corrupted rate is less than 40%, VMFhas the advantage over PRVM. In terms of NCD, if the corrupted rate is less than 10%,VMF has the advantage over PRVM; otherwise, PRVM has the advantage over VMF.The computational time of PRVM is less than the same scale moving windows VMF. 6. Discrepancy measure Progressive Reduced order Half Switch vector Median filter—DPRHSM Signal pixels are gotten based on DM, and all pixels are processed by VMF but onlyto use signal pixels. DM has very good effect on searching noisy pixels which threechannels have been corrupted simultaneously. So in the corrupted pixels detectedprocesses, this sort of noise will be eliminated. Then signal pixels are processed byPRVM. The PRVM has good effect on filtering single-channel corrupted pixels,because there is the idea of scalar median in it, but its output is always one of the inputvectors. The performance of the proposed method is compared with AEM, SMF,MMVM, PSM, FEVM, VMF, VDF, MMFT and PRVM in terms of objectivecriteria—PSNR, MAE, NCD and NMSE. The noise model is MODEL1, the corruptedrate ranges from 0 to 80%. If the corrupted rate is less than 40%, PSM has a littleadvantage over other methods; otherwise, DPRHSM outperforms others in terms of allobjective criteria used. When the noise model is MODEL2-1, SMF, PRVM andDPRHSM outperforms other methods. In order to deduce a quantitative compellentconclusion, an aggregate analysis method of objective criteria—PSNR, MAE, MSE,NCD, NMAE and NMSE is proposed to evaluate effectiveness of filtersaforementioned. The aggregate analysis result (two noise model, 5 images) shows that 9of 10 optimal values of its evaluating indicator are obtained by DPRHSM, and only 1 of10 is gotten by other methods. 7. Recognition of Near Bank Scene Boundary between Water Surface and Bank Based on Fractal Dimension Fractal features are applied lonely in many image analyses, and then their results are...