SEARCH RESULT

Turkish Sign Language (TSL), which is a mother tongue of hearing-impaired individuals, is a natural language. Expression patterns used in this language are carried out within the settings of the rules of the language. In this study, a textual interpretation system from Turkish to TSL was developed. Within this context, a corpus consists of 230 sentences was composed. Beside the sentences included in the corpus, also the interpretation of the sentences entered by the user interactively can be made. The rules of both languages were taken into consideration in the interpretation of the sentences. Firstly, Turkish sentences were parsed to words and then, morphological analyses of the sentences were performed withZe mberek. As a result of the morphological analyses, root/stem of the words and the affixes attached to the words were determined. Some affixes are not considered as necessary in TSL. Therefore, the affixes which are not considered as necessary should be determined and ignored within the interpretation. Moreover, various rules were constructed for the transformation by considering the results of the morphological analysis rules and usage samples. Descriptions related to 489 signs and 6 non-manual signs were made in order to express the sentences included in the corpus in TSL. The number of the signs out of 489 were as follows: 81 static, 408 dynamic, 334 single, 155 repetitive, 6 sign union and 2 word combination. 6 non-manual signs were as fol l ows; baş önde (head ahead)" bö", baş yukarı da (he ad up) "by", kaş yükseltme (eyebrow raising) "ky", kaş i ndi rme (e yebrow l owe ri ng) "ki", past aspect "di" (past tense suffix) and continuous aspect "yor" (present continuous tense suffix). Usage numbers of the nonmanual signs were as follows: 66 "bö", 12 "by", 38 "ky", 40 "ki", 94 "di" and 31 "yor".

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Mehmet Fatih Karaca Şafak Bayır

Optical Character Recognition (OCR) is the extraction of letters, numbers, punctuation marks, shortly the texts, from the images in the digital settings. As a result of this process, electronic images are converted into editable texts. In this study, the performances of character recognition techniques on upper-case letters, lower-case letters and numbers were evaluated in terms of various features. At first, training images with 10 typefaces were generated. Then, testing images were obtained by adding extra 10 typefaces into the system. The training images were formed in 36pt font size. However, the testing images were also formed in 12, 36 and 60pt font sizes in order to see the effects on OCR. Firstly, images were parsed into the characters, then, these images were stretched or squeezed into 100x80, 50x40, 25x20 and 5x4 pixels. Then, thresholding was applied to image files and each character was expressed as vectors having pixel value either 1 or 0. For the OCR process, 3 algorithms were used; k-Nearest Neighbors (kNN), S upport Vector Machines (S VM) and Artificial Neural Network (ANN). Eucli dean Di stance, Inner Product and Cosine Similarity were used for the measurement of similarity in kNN. The following results were obtained when the results are evaluated in terms of average means; the best classification was realized with ANN and the least with Inner Product (k=3) in upper-case letters, the best classification was realized with ANN and the least with Inner Product (k=10) in lower-case letters and the best classification was realized with ANN and the least with Inner Product (k=10) in numbers. The best classification was realized with 36pt and the least with 12pt i n upper-case and lowercase letters; and the best classification was realized with 60pt and the least with 12pt in numbers.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Mehmet Fatih Karaca Şafak Bayır