Table 3.

Performance comparison of TrOCR-ctx, TrOCR, Abinet, and PP-OCRv2 models on UoS_Data_Rescue, CORD, SROIE, PubTabNet, and ICDAR 2015 datasets using segmented text lines. Evaluation metrics include Rouge-L, Word Error Rate (WER), Character Error Rate (CER), Exact Match, and F1-scores at both character and token levels

OCR Model	Rouge-L	WER	CER	EM	F1-score (Char)	F1-score (Token)
	UoS_Data_Rescue
TrOCR	0.849	0.055	0.047	0.825	0.963	0.945
TrOCR-ctx	0.857	0.049	0.035	0.847	0.966	0.951
Abinet	0.545	0.557	0.346	0.432	0.681	0.449
PP-OCRv2	0.812	0.348	0.178	0.646	0.825	0.666
	CORD
TrOCR	0.898	0.168	0.056	0.802	0.946	0.834
TrOCR-ctx	0.957	0.034	0.016	0.833	0.985	0.986
Abinet	0.520	0.356	0.304	0.574	0.710	0.644
PP-OCRv2	0.789	0.114	0.144	0.746	0.897	0.886
	SROIE
TrOCR	0.919	0.053	0.044	0.830	0.984	0.947
TrOCR-ctx	0.940	0.033	0.014	0.849	0.988	0.967
Abinet	0.872	0.629	0.497	0.301	0.507	0.381
PP-OCRv2	0.882	0.432	0.235	0.493	0.777	0.577
	PubTabNet
TrOCR	0.878	0.141	0.069	0.748	0.940	0.859
TrOCR-ctx	0.913	0.091	0.067	0.789	0.965	0.909
Abinet	0.315	0.813	0.450	0.153	0.598	0.197
PP-OCRv2	0.833	0.131	0.097	0.700	0.915	0.877
	ICDAR 2015
TrOCR	0.777	0.245	0.102	0.744	0.904	0.750
TrOCR-ctx	0.776	0.250	0.102	0.749	0.905	0.755
Abinet	0.151	0.741	0.624	0.259	0.436	0.259
PP-OCRv2	0.665	0.374	0.178	0.626	0.831	0.627