A Research On Accuracy And Speed Of The OCR Systems In Specific Fields

During the past fifty years, people have gotten brilliant achievementsin OCR area. The development in OCR theory and technology make itbecoming possible to meet the needs for practical OCR products in finance,postal automation, news publishing, office automation fields, etc. Though,the automation of bank check processing systems and information retrievalonly have been studied for few years, the strong motive force of marketand society's needs make it becoming one of the hotspot in OCR area, andthen speed the development of the OCR theory and technology.The research of present thesis involves related technologies on theaccuracy and speed of OCR systems. The following are the resultsachieved in this dissertation:1．In pattern recognition, the number and quality of learning patternsis of crucial importance. When the number and quality of learning patternsare limited, error occurs in the presumed distribution of patterns and theprecision of whole recognition system decreases. A new pattern generationmethod is proposed which contributes to improvement of the performanceof a handwritten Chinese character recognition system. By using thispattern generation technique, we increase the number of learning patternsby using transform method with cosine function. Patterns generatedartificially this way are then selected using pattern selection method andthe patterns unsuitable for learning are discarded. 2．A new feature extraction method contributing to improvement ofthe performance of a handwritten Chinese character recognition system.By using enhanced weighted dynamic meshes based on nonlinearnormalization, this method not only avoids the zigzags and otherundesirable side effects introduced in the original Yamada et al.'s nonlinearnormalization method but also avoids additional feature normalizationprocess in dynamic mesh method.3．PCA(principal components analysis) and LDA(linear discriminantanalysis) based feature dimension reduction for the handwritten Chinesecharacters is investigated. PCA and LDA can effectively reduce thedimension of the feature vectors without decreasing recognition ratemarkedly, while LDA can improve the performance of recognition withcertain dimension reduction rates.4．The recognition of hand-written Chinese characters usingMahalanobis distance is extensively utilized in bank cheque processingapplications. The Mahalanobis distance, defined by the innovation and itscovariance, is compared among several target character classes, and thecomputation is a time-consuming operation. We present an efficientcomputation for this process. The method described here can besummarized as an incremental, non-decreasing computation for theMahalanobis distance; if the incrementally computed value exceeds thethreshold then the computation is stopped. The elements of covariance andinnovation are only computed if they are used, and progressivity is themajor advantage of the method. This method is based upon thesquare-root-free Cholesky's factorization.5．OCR based image retrieval and image filtering technology isdiscussed, and how to make more efficient use of traditional OCRtechnology is investigated.