Study On The Joint Transform Correlation Object Recognition Technology

Optical pattern recognition has the advantages of superior high speed and parallel processing, which make it a very important way to realize pattern recognition practically and in real time. It has been widely studied and used in the fields of military target recognition, automation control and fingerprint recognition.In this paper, two methods of realizing optical pattern recognition, spatial matched filter correlation and joint transform correlation(JTC), were discussed first with the emphasis on the JTC. This paper analyzed the modulation and filter technique on joint power spectrum(JPS), which reduced the intensity and width of the strong dc component existing in the classical joint transform correlator, improved the recognition ability greatly.Second, based on the theory of synthetic discriminant function, several sub-reference images were synthesized to form the reference image base to realize rotation distortion-invariant recognition, which also solved the ill-problem of synthesizing a single reference image with the whole training distorted images and made the pose estimation easier.Third, the phase-encoding and amplitude modulation distortion-invariant JTC for multi-target recognition were described. The proposed JTC greatly reduces the intensity and width of the strong dc component, sharpens the correlation peak, eliminates cross-correlations among each input target, and the conjugate cross-correlation between the reference and input targets at the same time.In order to calculate the discriminant probability of the objects, a simple but self-explainable statistic function was presented to show where the true object and false object occurred in the unknown input image sequence. This function also bodied such a fact that when the inputs continuously judged as the true objects, the probability of the true input would be strengthened, and on the other hand when some inputs were judged as the false objects continuously, the probability of the true input would be weakened.Based on the above theories, two softwares were developed for continuously shot input images and random shot input images. The softwares are with the ability of recognizing true and false objects, estimating the pose of the true objects andcalculating the recognition probability of both true object and false object in the input image sequence.Finally, the automatic opto-electronical hybrid joint transform correlator was set up, which produced a quite good recognition result for both binarized image and complicated grayscale-level images. This paper also discussed the effects of LCTV on JPS. A method of low-pass filter and appropriate off-focus was used to weaken the side-effect of the LCTV.