Works only for single word

[githubharald]

Works only for single word

[githubharald]

Hi,

yes, that's correct. The reason is not the number of words, but the size of the input image.
The input image is downsized to 128x32px, so if it contains a long sentence you won't be able to read the text on it anymore.
However, groups of short words should work fine, e.g. "I go home".

For more information see section 2 of this article: https://towardsdatascience.com/27648fb18519

[IamDixit]

If my image size is fixed to 128x32 for each input, I guess then multiple words will be read properly. Will this approach work?

…

On Wed, 19 Sep 2018, 15:15 Harald Scheidl, ***@***.***> wrote: Hi, yes, that's correct. The reason is not the number of words, but the size of the input image. The input image is downsized to 128x32px, so if it contains a long sentence you won't be able to read the text on it anymore. However, groups of short words should work fine, e.g. "I go home". For more information see section 2 of this article: https://towardsdatascience.com/27648fb18519 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#9 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AXV_7k6qcts36T4vixRDiSs0vQRHtmdgks5uchIogaJpZM4WvvTj> .

[githubharald]

There is no need to resize your input to 128x32, this is done automatically (also taking care that no distortion happens). You just have to take care that in your image, when downsized, the text is still large enough to be recognized.
I recommend either approach 2.1 or 2.2 of the linked article.

[IamDixit]

Thank you so much for the guidance

…

On Wed, 19 Sep 2018, 15:38 Harald Scheidl, ***@***.***> wrote: There is no need to resize your input to 128x32, this is done automatically (also taking care that no distortion happens). You just have to take care that in your image, when downsized, the text is still large enough to be recognized. I recommend either approach 2.1 or 2.2 of the linked article. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#9 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AXV_7mhdqmfPSV09Mn93dC9bj3hrBfvkks5uchetgaJpZM4WvvTj> .

[githubharald]

A small illustration:

• Top: your image containing some words.
• Left: downsize it and feed complete image to model.
• Right: apply word-segmentation (only one word shown). Feed each word individually to the model.

As you can see, the images of the segmented words contain much bigger text, which can easily be recognized by the model.

[IamDixit]

One more thing I have observed that for alphanumeric word(13BCE1004) the model gives poor result. Can you please tell me the reason for it and possible ways to fix it?

…

On Wed, 19 Sep 2018, 15:41 Abhishek Dixit, ***@***.***> wrote: Thank you so much for the guidance On Wed, 19 Sep 2018, 15:38 Harald Scheidl, ***@***.***> wrote: > There is no need to resize your input to 128x32, this is done > automatically (also taking care that no distortion happens). You just have > to take care that in your image, when downsized, the text is still large > enough to be recognized. > I recommend either approach 2.1 or 2.2 of the linked article. > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#9 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AXV_7mhdqmfPSV09Mn93dC9bj3hrBfvkks5uchetgaJpZM4WvvTj> > . >

[IamDixit]

You are right, using word segmentation I am getting correct results.

…

On Wed, 19 Sep 2018, 15:52 Harald Scheidl, ***@***.***> wrote: A small illustration: - Top: your image containing some words. - Left: downsize it and feed complete image to model. - Right: apply word-segmentation (only one word shown). Feed each word individually to the model. As you can see, the images of the segmented words contain much bigger text, which can easily be recognized by the model. [image: downsize_imgs] <https://user-images.githubusercontent.com/15148095/45747427-1dc3d880-bc06-11e8-857b-905ade59a9f9.png> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#9 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AXV_7td5E4bT6jB1v6vaZwrvnUV5AIugks5uchrdgaJpZM4WvvTj> .

[githubharald]

Is the writing style for you alhpnum images the same as for your "normal" text?
Just to be sure that the bad results are not just because the writing looks entirely different.

If it's not the writing style: maybe the model learned some language properties, such that certain character combinations are more likely than others, e.g. "na" is more likely than "nx" in the IAM dataset which the model was trained on. But this is just an assumption, I didn't do any experiments to check if this is really the case.

[PartheshSoni]

How can I train your network for some other language like Hindi??