Difference between revisions of "P2PaLA"

From Transkribus Wiki
Jump to: navigation, search
Line 13: Line 13:
  
 
==Training==
 
==Training==
Currently all pre-trained models are publicly available for all users. In a later stage of the integration, models will be associated with collections and users as with the HTR in Transkribus.<br/>
+
If you are interested in training your own models, please send us an email (email@transkribus.eu) and we can enable the training interface for you.<br/>
If you have your own dataset for training and recognition, please send us an E-Mail (email@transkribus.eu), then we can train a model for you.<br/>
+
Trained models will be associated to the currently selected collection and the owner of the model is set to the user who has trained the model.<br/>
Please make sure to tag structure types on ''region level'' only and avoid overlapping between different regions. Also specify if baseline detection should be trained too (which may only make sense for larger datsets). For structure type recognition, a training set of about 50-100 pages should be enough to generate a decent model, depending of course on the complexity of your layouts. Please note, that the tool can only recognize structure types that are in any way ''visually'' or ''positionally'' distinguishable on a page. Also note that P2PaLA is currently ''not'' a production-ready tool, thus please don't expect 'perfect' results.
+
When tagging regions, avoid overlapping between different regions. Baseline training only makes sense if your dataset is large enough, i.e. at least 500-1000 corrected baselines.
 +
For structure type recognition, a training set of about 50-100 pages should be enough to generate a decent model, depending of course on the complexity of your layouts. Please note, that the tool can only recognize structure types that are in any way ''visually'' or ''positionally'' distinguishable on a page. Also note that P2PaLA is currently ''not'' a production-ready tool, thus please don't expect 'perfect' results.
  
 
====Training Parameters====
 
====Training Parameters====
 
[[P2PaLATrainParameters]]
 
[[P2PaLATrainParameters]]

Revision as of 11:26, 12 November 2019

P2PaLA is a layout analysis tool that recognizes structure types on region level and baselines from a page based on pre-trained models.
The tool was developed by Lorenzo Quirós Díaz at the UPVLC in Valencia, see https://github.com/lquirosd/P2PaLA for the full Open Source codebase.

Recognition

Currently, the recognition is integrated into the Transkribus expert client (TranskribusX) for pre-trained models.
In this process, the P2PaLA tool creates new text-regions from trained structure types and optionally also baselines contained in those regions.
The table shows detailed information on all available models.
The column "Structure types" shows the list of region types this model recognizes and the column "Baslelines" shows if this models was also trained to detect baselines.

Parameters
  • Rectify regions -> all regions will be simplified to the bounding box of the actual recognized shape
  • Min-Area -> Shapes with an *area* smaller than this fraction of the image *width* will be removed after the recogniton. Use this parameter to remove small "garbage" regions. Default = 0.01

Training

If you are interested in training your own models, please send us an email (email@transkribus.eu) and we can enable the training interface for you.
Trained models will be associated to the currently selected collection and the owner of the model is set to the user who has trained the model.
When tagging regions, avoid overlapping between different regions. Baseline training only makes sense if your dataset is large enough, i.e. at least 500-1000 corrected baselines. For structure type recognition, a training set of about 50-100 pages should be enough to generate a decent model, depending of course on the complexity of your layouts. Please note, that the tool can only recognize structure types that are in any way visually or positionally distinguishable on a page. Also note that P2PaLA is currently not a production-ready tool, thus please don't expect 'perfect' results.

Training Parameters

P2PaLATrainParameters