Vision-Based Method In Biologic Strip Analyse System

Gold test strip is a newly developed method for biochemistry test, and it isused in many areas, such as disease detection and safeguard detection for foods.To use this method, we have to compare the test strips with standard ones andmake descisions according to what we see. As the development of biochemistrymethods, a lot of new equipments with new strips emerge, with which theprecision of the results are improved and the Gold test strip method has developedinto a directly quantitative analyse ones when necessary.Strip analyse systems in the market are mainly based on image measuringsystem, they acquire the images with image sensor, extract common characters ofbiochemical changes on the strips, then use the statistic data to analyse the images.There is only one kind of instrument for one kind of testpaper. There are newtendencies that the testpaper analyse system(such as analyse system for blood andbody fluid) can utilize all the characters, make decision intelligently,and can beconnected to computer easily.In this paper, the author use the computer vision theory to analyze theClenbuterol(a synthesized β 2 adrenergic receptor stimulation agonist)gold teststrip. Fistly , the author analyse the noises of the test strip and remove them withproper methods, then pick up the interest areas with right means;secondly, extractthe color and texture characters, set up classification machine based on fuzzyK-clustering of different concentration to be used in undefined concentration stripclassification, and the result is good;thirdly, implement USB subsystem for datatransfer of the system, here we choose a cheap, flexible USB chip and completethe firmware,driver and data transfer program;finally, a computer platform usingthe algorithms mentioned in this paper is set up.The first part is about the general description of the biologic strip analysesystem and USB transportation implement. The author introduces system frame-work and dataflow of USB, chooses PDIUSBD12 according to the prices,credibilities, flexibilities, then make up the hardware. A suitable firmware isdeveloped according to handy USB firmware.The second part is about the preprocessing of testpaper image, image noiseremove and image segmentation. Here the images obtained with the hardware areanalysed, the results show that the most serious noise are the impulse spot noiseand some Gaussian noise. The author choose the median filter which is the bestfor impulse noise. In order choose the proper kind of median filter, a serial ofmedian filters are considered. The results turn out that the filtering performance ofclassic rectangle window median filter are better when the enlarge the window,while capabilities of preserving the details worse;times of rectangle windowmedian filtering can not improve PNSR distinctively, nor more times of filtering;the adaptive window median filter can deal with impulse noise of different intensein different areas ,but the false rate is a bit high;this paper utilize a linear thresholdadaptive median filter, it regard the maximin as the noise and eliminatethem ,calculate the average of the rest , compare the absolute value between theimage pixel and the average with a linear value, then decide wether export themedian of the window. The last method shows better filtering performance. Inorder to remove Gaussian noise, average filter and Gaussian filter are compared,but the Gaussian filter is chosen for its better performance. In the end, the noiseremoval results of gold test strips are shown.For the image segmentation, segmentation based on the border and 4 ordina-ry differential operators are introduced. As the test paper images are mentioned,the Otsu adaptive threshold district method is used, and it is used between simpletarget and background. The distance between two classes is redefined, and thedisperse rate are used, the discriminant for best threshold is calculated accordin-gly. That method is used in multi-peaks image segmentation, and good results areshown. In the end, interest areas picked up from the gold test strips are shown.The third part is about the testpaper feature extraction and patternclassification. Color feature is the global feature which is definitely defined andirrelative to rotation, shiftting, size. In this paper, the dimension of the image isreduced from 3-D to 2-D, which reduce the computing amout and complexity. Theaverage, square error, skewness ,energy of the color feature are extracted forreference, and the square error is eliminated according to the fact. The texturefeature is the quantitative description of the special information for the image, andthis paper utilize the co-occurrence matrix,and redefine a co-occurrence matrixwith the distance of 2. As the graylevel boundary of the paper is to large, theauthor consider the actual boundary and compress from 256×256 to 60×60。Atlast, the entropy W E,contrast WC ,symmetry W H(k=20) are selected for late use.As the color and texture feature of the testpaper in the feature dimensionhave multi-apexes and interleaving, a simple non-linear pattern recognitionmethod-K nearest neighbor which is appropriate for small scale samples isused here. The author introduce the theory of K nearest neighbor, as the distanceof classes is not considered, membership degree function is introduced to solvethe problem. The author sets up classification machine for testpapers of fivedifferent concentration, then check it, finds out that correct probability is amost88% and it satisfys the system.The last part is about the computer platform for biologic strip analyse system.In this part all the algorithms metioned above are used, and an USB transportmodule is implanted. The author considers the conditions of driver development,then develops the USB driver with Driverstudio and VC.