chinese font problem

[DLUTfangping]

some chinese fonts can not generate good samples(for example ,some word could not be generated),do you have some suggests to solve the problem .thank you in advance

[Belval]

Hi!

Please include the chinese words you were unable to print.

Thanks

[DLUTfangping]

Thank you for you replying!
If I use the font that you provided ,I can generate good sample but when I change the font(eg.use other .ttf file ) the result maybe not so good.

[Belval]

That would be because the font you used does not support all characters. I'll try to provide more fonts in the future.

In the meantime, I know https://github.com/JarveeLee/SynthText_Chinese_version/tree/master/data/fonts has a lot of choice but I cannot add them to this project over copyright infringements concerns.

[DLUTfangping]

Thank you very much !!!

[liangshuang1993]

@DLUTfangping , have you tried training using dataset generated by this? I'm using crnn to train, but the result is not good.

[Belval]

@liangshuang1993 I can't say for Chinese, but I got decent result in English when using lowercase only (lowercase plus uppercase was a challenge). Also, while I don't know which implementation of CRNN you used, mine takes a long time to train (+50 hours on GTX 1080Ti) so it's very normal if the initial performance is very poor.

[liangshuang1993]

Hi @Belval , thanks for you answer.

Maybe I generated training data wrongly.

First I generated one dataset with word length is 5, using Gaussian Noise as background. the performance is good on training dataset and validation dataset, bad on some real pictures.

Then I generated another dataset with word length is 8, using given pictures as background. And I trained crnn on the whole dataset(dataset1 and dataset2). I have trained 7000 epochs, but the training dataset accuracy is still 0.5590. Strange thing is when I did some test on training data, I found it can barely recognize the word. So is this means I must have same length dataset? Thanks a lot.

By the way, each dataset has 500,000 pictures, containing English, Chinese and number.(They may appear in the same picture).

[Belval]

@liangshanghuang1993 That is indeed a rather ambitious idea to learn both English and Chinese. Most implementation I know only do one.

But yes, same word count is required. The original author even went as far as only recognizing single words instead of multi-word sentences.

[liangshuang1993]

OK. Thanks a lot!!

[yts19871111]

您好@Belval，谢谢您的回答。

也许我错误地生成了训练数据。

首先，我以高斯噪声为背景，生成了一个字长为5的数据集。在训练数据集和验证数据集上表现良好，在某些真实图片上表现不佳。

然后，使用给定图片作为背景，生成了另一个字长为8的数据集。我对整个数据集（数据集1和数据集2）进行了crnn训练。我已经训练了7000个时期，但是训练数据集的准确性仍然是0.5590。奇怪的是，当我对训练数据进行一些测试时，我发现它几乎无法识别该词。那么这是否意味着我必须具有相同长度的数据集？非常感谢。

顺便说一下，每个数据集都有500,000张图片，包含英语，中文和数字（它们可能出现在同一张图片中）。

你好，请问这个问题你解决了吗，我也遇到这个问题了，在真实样本上的效果很差