A Survey on Breaking Technique of Text-Based CAPTCHA

Table 3

Comparison of typical methods based on segmentation for breaking adherent CAPTCHA.

Example	Source	Success rate	Reference	Breaking method	Year
	Google,; Yahoo	4.89%–66.2%	[2]	Segmentation: width; Recognition: CNN	2004
	Microsoft; Google; Yahoo	61%; 8.7%; 25.9%	[4]	Segmentation: color filling and projection; Recognition: CNN	2008
	Hotmail	40%	[5]	Segmentation: change width; Recognition: SVM; Post-processing: DP search	2009
	MSN; Yahoo	18%; 45%	[6]	Segmentation: projection and central	2010
	Megaupload	78%	[36]	Segmentation: color filling; Combination: nonredundancy; Recognition: CNN	2010
	reCAPT-CHA Google	33%; 46.75%	[38]	Segmentation: character structure feature; Recognition: CNN	2011
	Yahoo	54.7%	[44]	Segmentation: projection and character feature; Recognition: OCR	2012
	Yahoo	36%–89%	[41]	Segmentation: color filling; Combination: redundancy; Recognition: CNN; Postprocessing: DFS	2013
	Microsoft	5.56%	[60]	Different width/location segmenting and template matching	2015
		57.05%
	reCAPT-CHA	40.4%–94.3%	[61]	Segmentation: trichromatic code	2015
				Recognition: SVM
	Yahoo	57.3%–76.7%	[7]	Edge and fuzzy logic segmentation and recognition	2015
	Microsoft	5%–77.2%	[42]	Segmentation: Log-Gabor filter; Combination: redundancy; Recognition: KNN; Postprocessing: DP search	2016
	MSN	27.1%–53.2%	[48]	Segmentation: different width; Recognition: BPNN	2016

Note. CNN: convolutional neural network, DP: dynamic programming, OCR: optical character recognition, DFS: depth first search, KNN: -nearest neighbor, BPNN: back-propagation neural network.