TextBoxGan: G enerating text boxes to train OCRs with a GAN
Автор: Theodo Data & AI
Загружено: 2021-04-08
Просмотров: 133
OCRs require a lot of labelled data to train. While there are a few synthetic datasets, generated with traditional methods (no AI), TextBoxGan is (to our knowledge) the first effort to generate text boxes with a GAN. Such approach offers different advantages: firstly, the generated text boxes are not constrained by a pre-defined font. Secondly, the style of the text boxes can be controlled to generate images from a specific distribution, which is useful for use-cases where your OCR has to read images that resemble each other.
During this talk, we will discuss :
(i) the network which generates text boxes from input words,
(ii) the datasets used to ensure our model performs well on every character, and (iii) the results obtained."
https://github.com/NoAchache/TextBoxGan
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
