Research On Localization And Recognition Of Billet Character In Complex Production Line Scene

Character localization is a crucial procedure in the optical character recognitionsystem. The results of locating process influence directly the following character trackingand recognition effect. However, there is almost no unified localization approach undervarious illumination conditions. Some theoretical issues and key technique of characterlocalization procedure should be researched and solved deeply in complex conditions,because it will promote the improvement of the intelligent production and the level ofmanagement in many of our industries, which brings huge development of potentialeconomic.This dissertation is designed in the background of billet characters automationrecognition of WUGANG heavy rail production line. Some localization methods of billetcharacter were analyzed from complex illumination scene. Through experimentalanalysis, it can be found that these conventional localization methods are difficult tolocate the character regions accurately from complex scene, especially the billet characterregions in complex illumination scene. In order to solve the scientific problem, thisdissertation mainly describes the following key technologies: complex backgroundsuppression, recursive segmentation of billet character, measuring and extracting billetcharacter.Firstly, an algorithm combining mean shift and connected domain analysis isproposed to restrain the complex background. A kernel function and a window functionare proposed in the Mean Shift filter algorithm with effective dipartite degree and strongdescription to restrain the background in complex illumination conditions. And then, thecombination of mean shift and connected domain analysis is adopted to separate area ofinterest and non-interest area effectively according to the regional features of thecharacter connected domain. This work provides the foundation for subsequent charactersegmentation processing.Secondly, a recursive segmentation algorithm based on separation evidence andseparation measure is proposed to segment the character images in complex illumination scene. The recursive segmentation algorithm is designed with criterion of maximumbetween-cluster variance. Billet images will be segmented step by step. The terminalcondition of recursion segmentation depends on the detection results of separationevidence and separation measure of the billet character. Vision invariant andquasi-invariant are integrated into separation evidence. The effective separation evidencecan be used to extract character local region in the billet character images of sequentialsegmentation. Then the measured data will be established to measure the result quality ofextracting region by projection evidence. Finally, we can decide which stage ofsegmented image is the terminal result (The complete information contained insegmented image) by comparing the measured data, and output character localizationinformation after accurate measurement.Experiments show that the proposed method can complete the localization preciselyin a wide variety of scenes. It has a certain degree of stability and adaptability in complexillumination conditions. After billet characters were located in complex scenes by theproposed method, the work of billet character recognition (segmentation, recognition)can be carried out smoothly.