Research On Hybrid Opto-electronic Correlation Pattern Recognition Algorithm

Along with the progress of the society and development of science, militaryand space technology, operation of large amounts of data for image recognitionand progressing in high speed is greatly demanded, which make the moderndigital electronic computer couldn't adapt to it. Hybrid opto-electronic correlationpattern recognition system is one of effective scheme to meet this need. Itincorporates the high-speed parallelism of an optical processing system and theflexibility and programmability of an electronic computer, makes best use of theadvantages and bypass the disadvantages of both, which solve the problem ofoperation speed for image recognition and progressing that pure electroniccomputer and optical processing system couldn't solve. Thus the hybridopto-electronic correlation pattern recognition has become one of the most activeresearch area in pattern recognition.In this paper, the development history of hybrid opto-electronic correlationpattern recognition is outlined, the basic theory of correlation pattern recognitionis reviewed. In addition, the principle of matched filter(MF) and joint transformcorrelator(JTC), which are two main methods for hybrid opto-electroniccorrelation pattern recognition, is analyzed and their implementation scheme ispresented. Then, the research status of hybrid opto-electronic correlation patternrecognition algorithm is summarized. Based on these basic research works,several correlation recognition algorithms are investigated and studied in detail.1.In pattern recognition, Wiener filter(WF) exhibits an optimal trade-offbetween noise tolerance and peak sharpness. In this paper, the application ofWiener filter is developed as follows:A rotation invariant Wiener filter is proposed. Traditional Wiener filter issensitive to image distortion. In this method, the reference term of WF's transferfunction is synthesized using the mean of training images. So the WF has themerit of peak sharpness, strong discrimination capability, noise tolerance androtation invariance at the same time. Furthermore, it is simple compared withother complicated composite filters which have similar capabilities. Although itdoesn't satisfy the equal correlation peak(ECP) criteria, it has no effect onrecognition because the correlation output dissimilarities between target andnon-target is large due to WF's outstanding discriminant ability. Simulationresults demonstrate the effectiveness of the proposed system. While, therecognition results are not same under different noisy environment. As far asnoise tolerance, the modified WF performs better under additive Gaussian whitenoise environment ; When it turns to the rotation invariance capability, itperforms better under non-overlapping high frequency color noise condition. A multi-targets recognition algorithm based on Wiener filter is proposed.Conventional WF is designed for one target recognition. In this paper, we achievemulti-targets recognition by modify the transfer function of WF. Recognition oftwo dissimilar targets simultaneously is considered and the simulation resultsshow the method has good performance under both noise and non-noiseenvironment. An modified WF under non-overlapping background noise is proposed. Inconventional WF method, the non-overlapping background noise is justconsidered as additive noise. In this paper, we modify the noise term of WF andpropose a method to extract non-overlapping background noise. Since thewindow function of the target image is considered, the modified WF outperformsnot only the normal noise estimation method, but also the conventional WF. 2.Quantized correlation output function to complexes of a quadruple(QFCQ)in the frequency domain is a new matched filtering method. It has sharperautocorrelation peak and much more suppressed non-target correlation signalsover those achieved by conventional inverse filter(IF), phase-only filter(POF) andbinary phase-only filter(BPOF). In addition, it has only one autocorrelation peakand so the location of the target is easier to detect compared with JTC basedtechniques. In this paper, we apply the QFCQ concept to distortion invariantpattern recognition. Synthetic discriminant function(SDF) filter is one of thebasic method for distortion invariant recognition. However, its main drawback isthe broad correlation peak and poor discrimination capability. In this paper, thecorrelation output function of SDF filter is quantized to complexes of aquadruple(QFCQ) in the frequency domain and an iterative algorithm is adopt toobtain sharp autocorrelation peak and satisfy the ECP criteria. Thus the distortioninvariant recognition is achieved. Two main distortions, rotation and scale...