How to improve training of custom models text detection?

[sossie07]

Hi all,

We have a pile of image files of assets that have been inspected in the field.

Most of the images have an Asset ID tag in the photo - for us this is aways a black tag with white characters in a vertical direction.

I have setup the AI builder to identify an photos with an Asset Tag in them, and this works well.

What I have done next is create a custom model to read the Asset Tag text, however I have found this to be unreliable - even though the layout of the asset tag is always the same.

Sometimes when I go and train the model it will pick up a "1" as a "7" or it will randomly insert a space " " into the text, which is not actually there.

I know I can add more images to the model, but that doesn't actually correct the incorrectly analysed text. Is there a way to improve the training? - e.g. tell the model what it should have actually read?

[Antrod]

Hi @sossie07 ,

Thanks for reaching out!

You can improve a model to better recognize fields in a document but there's no way to improve how the model can recognize characters. For example, if the model recognizes "7" instead of "1", it can't be improved with retraining.

Bad recognition often happens on documents with lower quality or when there are punctuations/characters very close to other characters or when font doesn't allow to differentiate some characters.

You can send us examples where this issue appears at aihelpen@microsoft.com and we can forward them our AI team, it could help improving future versions.