بازشناسي برخط و بدون محدوديت دست نوشته فارسي

چكيده انگليسي :

Online Unconstrained Farsi Handwritten Recognition Masood Faraki m faraki@ec iut ac ir Date of Submission 2009 4 22 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Maziar Palhang palhang@cc iut ac ir AbstractOnline handwritten recognition is a task of pattern recognition and artificial intelligence and has always beenin the special attention of researchers The difficulty of this field is doubled when considering languages suchas Farsi and Arabic since characters appear differently based on their locations in the words The analysis ofFarsi or Arabic scripts is further complicated in comparison to Latin script due to obligatory delayed stokesthat are placed above or below the majority of letters The input of an online handwritten recognition systemis a stylus and sensitive tablets like PDAs Few researches have been done in the field of online Farsi orArabic handwritten recognition Some of them are limited to the recognition of isolated characters or specialforms of writing of words are considered along with the use of a predetermined lexicon In the current project we have implemented an online Farsi handwriting recognition system in which theletters in the words can have any length and permutation Considering the lexicon has the advantage ofreducing the search space for classification and increasing the recognition rate In the current design we haveachieved these two advantages from a different view We have extracted path specification and relateddelayed stroke s like Noghte Donoghte for all Farsi characters in the design time In addition aninstance of a left to right Hidden Markov Model is considered for each character and delayed stroke exceptNoghte In the run time recognition process is done for each subword separately After delivery ofhandwritten points each body of a subword and its delayed stroke s are put in order based on a fuzzyinference system Then a string of delayed stroke names is constructed The main idea to generate candidatesubwords is based on path specification of characters and simultaneously matching their delayed strokes withthe pre recognized delayed stroke string Candidate generation is based on a proposed oversegmentationalgorithm The oversegmentation algorithm is based on some special points of Farsi handwritten charactersappearing in words After oversegmenting the input handwritten subword a graph subword graph isconstructed from the output segments of the oversegmentation algorithm At each node of the subword graph some useful information such as path specifications and feature vectors from the connected next nodes aresaved Final classification is performed by selecting maximum relative normalized probability of subwordcandidates The relative normalized probability of a subword is the sum of the probabilities of its charactersdivided by the number of the characters The probability of a character is computed by the extracted featuresand its corresponding hidden Markov model Due to the lack of a standard training and or testing onlineFarsi Arabic handwritten recognition data set we created our own set including 200 words written by 9people We have presented our result in terms of two metrics word recognition rate and subword recognitionrate The recognition results indicate a good precision in both writer dependent and writer independentexperiments Key WordsOnline recognition handwritten pattern recognition artificial intelligence hidden Markov model writer dependent writer independent

فركي، مسعود

بازشناسي برخط و بدون محدوديت دست نوشته فارسي

بازشناسي الگو , هوش مصنوعي , مدل مخفي ماركوف , وابسته به نويسنده , مستقل از نويسنده