The Research Of The Distortion-invariant Optical Pattern Recognition System

Optical Pattern Recognition(OPR) and optic-electricity-data-process increasingly exert a strong fascination upon researchers for its high operation speed huge resolution and low power consumption. For this reason ,researchers of varies nations court for the OPR application with the feature of real-time freedom high recognition possibility and low false alarm. However, practical OPR system is still in the procession, for there exists several complicated and key technologies. The author has threw a lots of times and energy into the research of OPR, including the constructions of hard ware and software algorithm and its character, and now taken a substantial progress. The key fruits is as follows:1 In the image-preprocessing, the optimal edge-extraction algorithm was used to increase the correlation S/N and auto-thresholding technique was introduced to decrease the affection of illumination distortion on the recognition possibility, and the bottle-neck of image-preprocessing was delivered to the DSP which is on the Genesis card instead of CPU.2, The optimal designed Digital Matched Spatial Filter(DMSF) was used to perform real-time optical correlation, which not only has the merit of large storage and renewability as well as common digital filter, but also has the virtue of high correlation S/N which the holographic optical filer owns. 3 In order to ensure high enough correlation S/N to any distortion version of recognized object, the distortion parameters tolerance research to 3D object in OPR has been done. Having studied the data tables, we got the relation between optical correlation S/N and distortion parameter (scale rotation azimuth) and then constructed the proper RFL, and so the puzzle of distortion of 3D objects in practical OPR system is now resolved .4 In the post processing of correlation signal, in order to improved the recognition possibility to the distorted multi-object, we increase the amount of information of ROI and choose the optimal numbers of neurons. So the ROI could include enough feature but does not make the CPU burdened too much. On the other hand, a new method of selecting ROI is realized, as a result, recognition possibility was increased and the number of samples for ANN trainning decreased.5 A multi-thread software cooperated with SMP (2 CPU) hardware performs the image-preprocessing correlation and post-processing synchronously, thus an automatic working module is realized.6 Systematic designed software, including grabing sarnies constructing RFL diging ROI, and trainning ANN offline and inline modules, has passed the practical tests which indicates that the OPR could reach the recogniton possibility above 85% and decrease the false alarm in the environment of high noise low contrast and multi-objects. And these are proved stable and repeatable in statistics, and so an advanced progress toward practical systems is achieved.7, Moreover, there discussed how to write a systematical migrationable easily maintainable software, and how to choose Operation System and developing tools. Because it comes from "software engineering" and integrated with the author's many years experience, it is universal and could be referenced and debuged by many other researchers.