Model very sensitive on PNG input

[junyizhao04]

I tried multiple size and source (photo from screen, paper, screenshot etc) and attempted to run it in Calamari OCR using the given model. However, the model is very sensitive towards input and only around 5% works. What is the expected PNG input size?

[andbue]

I don't think the input size matters that much. Please note that calamari is line based, it does not include code or models for segmentation tasks. As long as your images is cropped nicely around a single line or you provide coordinates via PAGE XML, it should work.

[junyizhao04]

I don't think the input size matters that much. Please note that calamari is line based, it does not include code or models for segmentation tasks. As long as your images is cropped nicely around a single line or you provide coordinates via PAGE XML, it should work.

then I believe image like this may work, but it does not.

[andbue]

The uw3 dataset the model uw3-modern-english was trained on contains only binarised data, therefore the model struggles with colour or grayscale images. If you convert your image to monochrome, e.g. via convert online.png -monochrome online.bin.png, it's recognised perfectly.

[bertsky]

You can see what a model expects by lookin at its data.input_channels and data.pre_proc.processors parameters (in the JSON file), or predictor.data.params.input_channels and predictor.data.params.pre_proc.processors (at runtime). If there is just 1 channel and the first processor is a CenterNormalizerProcessor, then this model expects binarization.

Besides the simple method mentioned by @andbue there are very sophisticated algorithms and models for binarization in OCR.

I have not seen public models for Calamari trained on grayscale or colour yet.

Can we close @junyizhao04?