| The technology of point target recognition and detection in complex infraredimage sequences is a difficult subject in the field of image processing and targetrecognition in recent years.Due to the long distance between infrared image sensor and the point target, itlooks like a point source and occupies only one pixel in the image obtained by thesensor without any shape characteristics. The SNR of the image is very low and thetarget is overflowed by noise, so the detection becomes quite difficult. Combined withmany detection algorithms given by the former, the applications of wavelet transformin point target detection are mainly discussed in this paper. The main content of thisdissertation are:The model of infrared image sequences containing the small targets is introduced.The features of the targets and the background and the noises are analyzed. Twopre-processing methods are proposed for the analyses of the small image sequences.Some characteristic of the targets which are relatively stable are proposed. The featureof the small targets in the frequency -domain is researched,using the characteristics ofthe small targets for identifying the true small targets. The process of single frameimage of the infrared image sequences which includes background segmentingsuppressing and enhancing is proposed. The method of the targets is also researched.A new filter is proposed which is a dual meaning filtering. Comparing to the otherfilter, the filter have some advantages such as high SNR(signal to noise ratio) andcapability of self-adaptive. Based the assumption that is the displacement of the smalltargets among image sequences are very small, a method energy accumulation isapplied for the image sequences after filtering. The theoretic is proved for the gain ofSNR. The conclusion of the methods of the small targets segmenting is done; a newmethod for the small targets segmenting is proposed which is maximums visualcontrast. The method need not the management of the human, and can work well onthe situations containing all kind of SNR. The feature of the temporal-domain of theimage sequences is analyzed. The character of III the targets in the temporal-domainin the image sequences is observed, so a new method of target segmenting in temporaldomain is proposed. According to the multivariable in the probability, a new methodof combining spatial feature and temporal feature for segmentation is proposed, whichcan be realized by parallel scheme and save amount time, and the precision of thesegmentation is improve highly. The applications of wavelet transform in point targetdetection are mainly discussed in this paper. According to different correlation of backgroundsignal and noise between neighboring scales, the approach of point targets detection based onwavelet correlation algorithm has been tried and investigated With the time-frequent analysis andmulti-resolution analysis of wavelet transform. Using the adaptive threshold methods and thesegmentation algorithms, the point target can be separated from the background. This algorithmcan overcome the shortcoming that traditional filtering methods can not distinguish between thenoise and edge signals of the infrared image. Now we introduce the methods as following.1.Double mean-value filterThe main idea is design two windows named as micro-window andmacro-window. The micro-window is design to measure the ares grey levelcharacteristic where is put pixel as centrer. The macro-window is design to measurethe larger ares grey level characteristic and then compare the conclusions which drewfrom above. The process as following:1) Use a big template (area of template are bigger than objects) to filter original figure with average filter method to yield figure I1. Due to area between object and isolated little noise module is small, the component they take in template is much smaller than background does, when the results of mean value processing is close to smoothed background gray degree.2) Use a small template (area of template is equivalent to object) to filter original figure with average filter method to yield figure I2, when the results obtained in the background area is background smoothed result, while the size of object is not so different from template and the distribution of gray is well-proportioned in object area, so the filtered gray degree is more close to object.3) Calculate the absolute value of difference of two filtered figure above and figure I3 i.e.: , I3 = I2 ? I1The size of template will not effect too much on the above two processed resultsof background area and gray degree change is smaller, then counteract in effect can beobtained after subtraction; object area exists definite difference due to the two filteredresults but tones up after subtraction. This method is not only effective to "brightobject"but also applicable to "dark object".2. Max Segmentation of ContrastThis method takes use of max absolute contrast to obtain gate limit to segment. Firstly, absolute contrast is defined as below: c(a,b) = min( F(a) ? t , F(b ) ? t )In the formula F(a) is mean value of gray higher than gate limit t area.,andF(b)is mean value of gray lower than gate limit t area. Finally the best segmentationgate limit T can be obtained: T = Arg max(c(t)) 0≤t≤255This method is able to self-adapt to obtain gate limit, and some parameters is notnecessary to set or to calculate, so this method is used widely. As long as gray ofdegree of object and background exist some difference, it is able to implementsegmentation more accurately, and therefore it is an ideal method.3. Object Segmentation Based on Combination of Time and SpaceBackground restrain Enhance objectSequence of image Image syncretism Image segmentationTemporal filter morphologic expandSince temporal and Space is independence each other, in order to improve the speed ofprocession, we can adopt the method by parallel processing.1) The space procession: In order to correspond with the temporal procession, we suppressin background to frame image of t,t ?1,...t ? L , then carry on the energy storage to strengthenthe object, and restrict the gray value of the image in 0~1, and this image is the process result F1of t;2) The Temporal procession: Because when we filter on temporal we only consider oneimage element, and the method to per-image-element is same , we can adopt hardware to improvethe speed of the procession at this condition. To the filtered image, we also should consider thecondition of the object of superposition in inner frame, firstly, we change the filtered image's graylevel to 0~255, then expand gray level, at this moment the structure element's size often is 3x3,and then expanded gray level of image will change to 0~1for following operation, at this momentthis image is marked as the space procession result F2 of t:3) we combine the space and time by using the formula F(x,y) = F1(x,y)? F2(x,y)tostrengthen the object, in order to make it convenient for segmenting the object, and the gray levelof syncretized image is varied to 0~255, the image at this moment is marked as the filter ofcombine the space and time result F2 of t:it4) To image F, we attain the limit value through the formula T = Arg max(c(t)), and 0≤t≤255segment the object.From procession of image, we can know the result of segment after syncretized is obviousbetter than the one who is single segmented apart in single frame and only segmented in temporal4. Correlative algorithms on the basis of the waveletThe procedure block diagram of whole algorithm is as follows: Algorithm of Adaptive Original Image relevant filter threshold of image pretreatment in wavelet object at field detected ç›®æ ‡æ ‡æ³¨ point Label of objectRelevant filter algorithm on the basis of the area of waveletWe define W (m, n) as the dispersed wavelet vary coefficient of the signal on m ofyardstick, and choose the vary value on the coterminous yardsticks for relevant calculation. Wedefine that relevant value is Corr2(m,n), the square difference of noise in every of yardstick isδm. The algorithm has two important steps, one is calculating relevant quantityCorr2(m,n),and doing the normalization of energy, and then combine with the relevant value of W(m,n) todecide the choice of the vary value W (m , n ), the other is choosing the square difference of noisein every of yardstick as the standard of finishing iterative.The step of algorithm of the relevant filter on Multi-yardstick is are as follows:(1) get the relevant quantity of Corr2(m,n) between every yardstick and... |
PDF Full Text Request