is Layout analysis trainable and works?

[josef821]

hi. i want to train and use your Layout analysis. does it working? have you any output of that?
thanks.

[robertknight]

The layout analysis model is a non-functional experiment at present. The rough idea will be to implement something like https://arxiv.org/abs/2203.09638.

Depending on your needs, a solution in the interim may be to treat layout analysis as problem of detecting objects in images. There are various pre-trained models that exist for this, plus various resources for training your own. See https://huggingface.co/Oblix/yolov10m-doclaynet_ONNX_document-layout-analysis for example. The underlying inference engine that Ocrs uses can also run many of these models (including YOLO). There are some object detection examples here.