Robust Text Recognition for Uzbek Language in Natural Scene Images

Improvement of the end-to-end scene text recognition method for “text-to-speech” conversion

“…This design decision was very time-consuming and expensive in computation; however, it had an important benefit: by examining the nesting of forms, and the number of child and grandchild forms, it is easy to identify and recognize inverted text. 26 In this step, forms are grouped, essentially by nesting, into blobs. Blobs are built into text lines, and the text lines and regions are examined for set pitch and equivalent text.…”

“…After the cropping probabilities are tried, a best-first search of the resulting segmentation graph repairs cropped character components or pieces of characters that were divided into multiple CCAs in the given image. 26 At every stage in the best-first search, each new blob compound is classified, and the classifier outputs are again provided to the dictionary. The result of a word is the character string that had the most beneficial overall distance-based evaluation, following weighting corresponding to whether the word was in a dictionary or had a reasonable combination of punctuation around it.…”

“…Because the adaptive classifier may have discovered something too late to contribute to the top of the page, a second pass is made, in which words that were not identified fully enough are identified again. 26 In the last stage, fuzzy spaces are determined, and different hypotheses for the xheight are examined to place tiny cap text. This algorithm also includes recognized text sent to the TTS synthesizer for the Uzbek language.…”

“…The second result, named the rating, increases the normalized distance from the prototype by the entire outline length in the unknown character. 26 The total outline length for all characters in a word is constantly the same, and it helps ratings of characters inside a sentence to be meaningfully summed.…”

“…In specific systems, this part involves the calculation of the target prosody (pitch contour, phoneme durations), which is then imposed on the output speech. 26 The proposed scene text detection and recognition method act as preprocessing for an Uzbek language TTS system by preparing raw text data from given natural scene images, as shown in Fig. 12.…”