**5. Technical performance testing**

First of all, we started with the technical performance testing to test the technical aspects of the proposed handwritten CAPTCHA. A total of 500 different CAPTCHA images were generated for each language. Each of the 500 CAPTCHA images of the first 3 languages (English, French, and Spanish) were tested on 6 different OCRs, while the 500 Arabic CAPTCHA images were tested on the second and sixth OCRs. **Table 1** illustrates the testing results for the six different OCRs on each adopted language.

As shown in **Table 1**, the testing results were divided into four patterns [1]:


correctly recognized, 47.8% no text found in the CAPTCHA image, and 47.9% totally incorrect text recognition. Likewise, in the French language case, the six OCRs used in the experiments did not succeed in correctly recognizing any of the French CAPTCHA images, while only [7] 0.4% of the whole French CAPTCHA images were partially correctly recognized, and 53.8% of the images resulted in no text found, and the remained 43.8% of the images were incorrectly

Similarly, the experiments result for the Arabic language shows that the two used OCRs have failed to correctly recognize any of the Arabic CAPTCHA images. However, the two used OCRs were able to partially correctly recognize only 0.3% from the whole Arabic CAPTCHA images, whereas 53.3% of the images resulted in no text found and 46.4% of the images were

recognized.

incorrectly recognized.

**Pattern Correctly** 

**recognized**

First OCR English 5 54 431 10

Second OCR English 0 30 390 80

Third OCR English 0 65 250 185

Fourth OCR English 0 40 150 310

Fifth OCR English 0 0 115 385

Sixth OCR English 0 5 85 410

**Table 1.** The testing results for the six different OCRs on each adopted language.

**Partially correctly recognized**

Spanish 1 16 475 8 French 0 2 486 12

Spanish 0 24 403 73 French 0 18 398 84 Arabic 0 3 391 106

Spanish 0 47 277 176 French 0 23 201 276

Spanish 0 37 84 379 French 0 29 63 408

Spanish 0 0 81 419 French 0 0 97 403

Spanish 0 3 118 379 French 0 1 68 431 Arabic 0 0 73 427

**Incorrectly recognized**

http://dx.doi.org/10.5772/intechopen.72599

Innovative Multilingual CAPTCHA Based on Handwritten Characteristics

**No text found**

171

**4.** No text pattern found: it is when the OCR was not able to recognize the text or any character in the CAPTCHA image.

In the English CAPTCHA images, the six OCRs have failed to recognize the full text in 99% of the images, while only 1% was correctly recognized. Nevertheless, the 99% includes 6% partially correctly recognized patterns, 46% no text found in the CAPTCHA image, and 47% totally incorrect text recognition [1].

Moreover, the other languages testing outcomes resulted in a lower recognition percentage compared to the English one. In Spanish language case, all the six OCRs have failed to correctly recognize 99.97% of the Spanish CAPTCHA images; this 99.97% includes 4.23% partially


**Table 1.** The testing results for the six different OCRs on each adopted language.

OCR software supports 46 recognition languages and it is able to extract texts in any of these languages. It also can detect text written in more than one language in the same image or

NewOCR is a free online OCR service that we used as the sixth technique in our experiments. The NewOCR service is available in the following link: https://www.newocr.com/. This online service supports more than 100 recognition languages and different fonts supports. In addition, the NewOCR service works using Tesseract OCR engine which is considered the best accurate OCR engine available at this time. It also supports the low-resolution images and can

First of all, we started with the technical performance testing to test the technical aspects of the proposed handwritten CAPTCHA. A total of 500 different CAPTCHA images were generated for each language. Each of the 500 CAPTCHA images of the first 3 languages (English, French, and Spanish) were tested on 6 different OCRs, while the 500 Arabic CAPTCHA images were tested on the second and sixth OCRs. **Table 1** illustrates the testing results for the six different

**1.** Correctly recognized pattern: it is when all the text in the CAPTCHA image has been cor-

**2.** Partially correctly recognized pattern: it is when the OCR has recognized three or more

**3.** Incorrectly recognized pattern: it is when no characters have been correctly recognized in

**4.** No text pattern found: it is when the OCR was not able to recognize the text or any charac-

In the English CAPTCHA images, the six OCRs have failed to recognize the full text in 99% of the images, while only 1% was correctly recognized. Nevertheless, the 99% includes 6% partially correctly recognized patterns, 46% no text found in the CAPTCHA image, and 47% totally

Moreover, the other languages testing outcomes resulted in a lower recognition percentage compared to the English one. In Spanish language case, all the six OCRs have failed to correctly recognize 99.97% of the Spanish CAPTCHA images; this 99.97% includes 4.23% partially

As shown in **Table 1**, the testing results were divided into four patterns [1]:

document.

**4.6. Sixth OCR**

170 Multilingualism and Bilingualism

extract the text written in these images.

**5. Technical performance testing**

OCRs on each adopted language.

characters in the CAPTCHA text.

ter in the CAPTCHA image.

rectly recognized.

the CAPTCHA text.

incorrect text recognition [1].

correctly recognized, 47.8% no text found in the CAPTCHA image, and 47.9% totally incorrect text recognition. Likewise, in the French language case, the six OCRs used in the experiments did not succeed in correctly recognizing any of the French CAPTCHA images, while only [7] 0.4% of the whole French CAPTCHA images were partially correctly recognized, and 53.8% of the images resulted in no text found, and the remained 43.8% of the images were incorrectly recognized.

Similarly, the experiments result for the Arabic language shows that the two used OCRs have failed to correctly recognize any of the Arabic CAPTCHA images. However, the two used OCRs were able to partially correctly recognize only 0.3% from the whole Arabic CAPTCHA images, whereas 53.3% of the images resulted in no text found and 46.4% of the images were incorrectly recognized.
