Patrick Li

是的，windows下只保留小写或者大写就可以了，用set过滤也不错。不管也可以，不影响训练。

half a day using one titan xp. you may train several times to find the best results. BTW this model is for studying tensorflow api, it's too simple in applications...

thanks, i'm sorry to reply in time, you can ask any questions on this issue or send mail to [email protected].

In this project, every character is distributed in specific areas of the image, so CNN can combine localization, segmentation and recognition of multi-character text. References： https://vision.in.tum.de/_media/spezial/bib/stark-gcpr15.pdf

这个项目里，每个字符都分布在验证码图片中的固定一块区域内，使得CNN能够对每个固定区域中的字符进行识别。 可以参考：https://vision.in.tum.de/_media/spezial/bib/stark-gcpr15.pdf