2026-10-13T03:52:47Z https://www.scidb.cn/oai

10.57760/sciencedb.39384 2026-06-11T15:24:44Z

2026-06-11 A Tibetan Scene Text Dataset under Natural Degradation doi:10.57760/sciencedb.39384 en To support the study of Tibetan scene text recognition under real-world natural degradation, we construct TibNST, a Tibetan scene text dataset collected from plateau outdoor scenes. Different from datasets acquired under controlled conditions or generated by synthetic degradation, TibNST focuses on naturally degraded text images captured in real environments. The dataset covers multiple practical scenarios, including commercial advertisements, road and directional signs, street posters, public notices, and temple-related scenes. TibNST contains 2046 Tibetan scene text samples. Each text instance is manually verified and transcribed by annotators with Tibetan reading ability, and the corresponding text-region location and line-level transcription are provided to ensure annotation reliability. The dataset therefore reflects real degradation factors in plateau environments, such as illumination imbalance, blur, scattering, low contrast, edge erosion, and background interference, which can damage Tibetan structural cues and increase recognition difficulty. During data collection, TibNST is formed into two subsets according to the observed degree of natural degradation, namely Subset1 and Subset2. Representative samples are shown in Fig.~\ref{fig:subset_comparison}. Subset1 contains samples with relatively mild environmental interference and clearer structural cues, whereas Subset2 contains samples affected by stronger degradation, such as severe blur, illumination imbalance, scattering-induced detail attenuation, and complex background interference. The two subsets contain 1532 and 514 samples, respectively. As shown in, Subset1 and Subset2 exhibit broadly similar trends in text-length distribution and high-frequency character composition. This indicates that the performance gap between the two subsets is mainly associated with different degradation intensities rather than substantial shifts in textual attributes. Therefore, TibNST provides a reliable experimental basis for analyzing the robustness of Tibetan scene text recognizers under different real-world natural degradation conditions.   Tibetan; dataset; TibNST Lhundrub Gyaltsan PUBLIC https://creativecommons.org/licenses/by-nc/4.0/ dataset Science Data Bank