 Good. Then next up is Ari Veselainen from the University of Helsinki and you will be talking about fine-grained OCR document layout analysis for the handprinted books from the 18th century using transcribers accuracy and challenges. Okay. Yeah, thank you. So, okay, first of all, although that now I'm talking about OCR, it's not really only related to the OCR, it's also basically handwritten, but because of the data I'm using, I'm focusing on OCR. So, actually, I have been working on my master thesis, and as part of the master thesis I have implemented more like this kind of fine-grained layout analysis model for the OCR texts. And this is based on the mask RC and instance segmentations, and I think that during this morning we heard a bit more about what are these different segmentation techniques and actually this is very close what was described earlier that what would be in a future transcribers. And all the work what I have done is currently done outside the transcribers, however, I have used P2pala as an example how to create these layout analysis and I'm basically then describing how this has been done and what we're defining. So this is only part of my thesis work. First, a couple of words about them. The problem here is that most of the layout analysis solutions available today are really focusing on this kind of cost-grained solutions, and I wanted to understand what does it then mean if you want to implement this kind of fine-grained analysis. And cost-grained I mean that in practice the solutions are only able to provide a more like this kind of high level elements, like in OCR case we are talking about text regions, fake pictures, tables, etc. So I wanted to go a bit more to the details and really understanding also the semantic meaning of those elements and the specific questions I had was that okay what does it then mean to implement this kind of fine-grained segmentation tool and how it should be then done and what would be then the expected accuracy and then what would be then the main challenges on implementing something like that. And then as I think that everyone of us are always struggling that okay is this then adequate for the larger scale project so with if the accuracy is not enough then you can probably deal with some manual fixes, but if you want to do it for the larger project then obviously you need to have as accurate as possible. And then to basically the work has been done partially in cooperation with this critical edition of David Hume's history of England project where they are currently working on understanding how the books of David Hume's evolved during the between the different editions as David Hume was actually revisiting and changing the text during the time. Okay then first couple of words about this p2pala. So, I don't know how many of you have been working it, but it's a basically an optional, I would say the optional more like a model of transcripts. So it can be used. And basically it can be used for the, for the detecting text lines, text regions and also classifying those text regions. It basically implements this kind of new shaped convolutional neural network and originality is this type of CNN was implemented for biomedical more like area, however, it's actually applicable applicable for almost any kind of image segmentation topic. And then Yeah, and here's basically just a little bit about what kind of data I'm working with. So this is a typical page or page opening from the from the 18th century. So the printing during that time was based on this hand printing techniques. And this is also reflecting on the structure of the pages. So basically the main components of the page are these. You have the main body. So that's normally what we are then really interested. Then we have the headlines we have the direction lines here on top of the top of the page and then we have all sort of marching up margins. And it is important to more like separate all these and then so that then you would be then able to use the material for the study. And okay, here is an example that if you then want to to annotate the data that you basically then draw and separate these different elements and this is such to more like show that okay where do you have the food notes where do you have the marching earlier where do you have these these these all all these elements so very, very straightforward, at least in principle. And here, the annotation, I was using the transcript was because actually that's quite handy tool. And, and, however, I was not, I had all together okay here is 660 pages actually I had little more than 700 pages. But first I I just implemented this manually and annotated hundred pages and then use the transcript was to create the first model. And then I more like iterated this and fixing manually the errors and finally then I had something like 700 pages available for my model. And okay here. So here he is more like the first page or example I want to show that that, as you can see, I would say that this is almost perfect, perfect interpretation of the page. All the, all the page elements are correctly detected and also are correctly located. The only maybe as minor minor problem is that you have some overlap between these regions. Okay, then here's an example and as, as you can see, this is maybe not that good and here it's not this kind of normal content page. This is a table of content page and and therefore, first I thought that the problem is probably due to the, I don't have enough examples in my table of content pages in my, my, my training data. But it might be that that's not the case and currently I'm still working on understanding why the model is showing such results. And then here are a couple of other examples so although that it works quite nicely for the pages but then you have some, some issues that sometimes you have this overlapping area sometimes some more like regions are combined and and a mixture of these so accuracy is not necessarily quite as good as you would like to have. And here I tend to run through some of the, the metrics to understand that how well is it is working. So the classification overall, I think that pretty good results so it's exactly it's able to detect these elements correctly on every page. But then, if you use another measure that's called these interconnection or intersection over union type of which is basically measuring that how well the box or the mask or segments of the regions are then located. You already see a bit mixed results so for this text body and header and headings, extremely good results but then for these other page elements. There's definitely a lot of improvement required. Just a couple of words about the conclusion so definitely this kind of segmentation methods can be used for this layout analysis. One of the examples provides good tools for this data set labeling and annotations and and be to power can be used for this this kind of work. I think that originally I made a wrong assumption I was, I was thinking that that you need to have a very detailed level of more like these elements detection and that it's not necessarily, because it looks that the model is not able to accurately detect all these different and therefore some further study is working and I'm currently working on this so thank you.