Extract Text, Title, Paragraph, Image From A Image Document using Deep Learning.

[@tikam75007]

thx man, it is pleasure to walk through your project. No bug, no surprise, really easy to go from one step to another, so smooth

[@ShivShankarDutta1]

excellent explanation

[@kudaykumar1261]

Thank you so much its really help full.

[@naveenkartikthangavel7237]

Hi sir, I'm actually confused at the step where we try to number the text blocks (the place where it assigns the numbering in the pic from 0 to whatever available), how does the numbering works? We are trying to retrieve the title, abstract, authors from a provided journal paper. We don't know how to retrieve those, as numbering is confusing. It would be really helpful if you give us some suggestions! Thank you..

[@chinmaykalinkar4656]

Hi Naveen! The numbering is basically distinguishing between text, title, figures etc. By just numbering you can differentiate between the types, either you give them labels like (title, text, etc.) or just give them numbers. If you will check the class "Detectron2LayoutModel" and its parameters that are passed. you will understand if you will not pass labels, it will by default assign the numbers to differentiate between the types. Hope this helps! Thanks.

[@TheChaoticTranquil]

Any reference on how to train layoutParser on custom dataset? Another question, how do you compare layoutlmv3 with layoutParser? Is Layoutlmv3 able to create bounding boxes around textareas similar to this?

[@karndeepsingh]

Both of the model have different architecture and are used for different purpose. Layoutparser is used for extracting generic layout informations like text, paragraphs, images etc. LayoutLM models used from extracting keyword information from images or pdfs using language model and layout information

[@kibtiachowdhury6011]

Hi Sir. This video is really helpful. It works. I have faced some problems. Some text block does not bounded perfectly. Some boxes shape is smaller than the real text . How can I get same to same text block size for each text block? How can I set segment_image?

[@karndeepsingh]

You may face little error while getting those boxes marked over the image and this is bcz of model error. You can increase the quality of image and try

[@kibtiachowdhury6011]

@@karndeepsingh Sir, I want to get paragraph and title from every pages of pdf. All pages does not work same config value and padding value. I had to changed for every pages. I want to get all the pages text at once, without changing it every time. How can I solve this?

[@RahulParmar-ld4ut]

Hi Karndeep, how to finetune this model on two classes: text and table?

[@karndeepsingh]

You can train object detection model with the dataset. Also, you can fine tune LiLT models with the dataset to detect the classes you need.

[@wilianuhlmann5284]

How do I create my own training? I have this need because I wanted to train Brazilian newspapers (Language Portuguese BR) and with this training base in English it is not performing well. Can you help me?

[@EhsanIrshad]

What about line segmentation?

[@nageswarreddy8626]

Please make a video on installing detectron2...No proper information available anywhere...it would be really helpful 🙏

[@karndeepsingh]

Check with there github page and raise the issue.

[@ipopa1995]

As specifically on windows.

[@karndeepsingh]

I have a video on Object Detection using Detectron 2. Please check the video on channel, you see how detectron 2 can be installed.

[@varunpusarla]

How can we change the color map ?

[@karndeepsingh]

You can check in the arguments

[@Moniragu]

how to download excel format

[@yuvrajjadhav1506]

Hi karandeep can you give me training to extract data from image and pdf

[@ribhavojha3638]

hello. I was trying to pip install layoutparser in coda env miniforge, but it just doesn't not work. Any idea why

[@karndeepsingh]

You can try to install layoutparser from official site

[@marouanezaid4066]

hello, actually im working on a model that extract automatically information from scientific articles (like title of the article, authors, date, published on ....) do you have an idea on how i can do it ? what should i use to achieve it ?

[@karndeepsingh]

Try using object detection models

[@marouanezaid4066]

@@karndeepsingh ok thanks, i will try

[@marouanezaid4066]

@@karndeepsingh and i have a question plw, can detecting object of a scientific article (like detecting title or authors) help extracting it automatically ?

[@20-reddyprakashg10]

Hi sir can we detect and extract only photo(user image) from document.

[@karndeepsingh]

Yes! You can!

[@20-reddyprakashg10]

Can I know how

[@sushruthbhat5727]

Hey, really helpful video. But is there a way I can extract only the images using this library?

[@karndeepsingh]

Check the documentation!

[@ytiam-unata]

Awsome

[@DANstudiosable]

Nice tutorial. Thanks alot! Is there any way where we can associate titles with their corresponding paragraph? Or at least print a paragraph by selecting their respective title. Example: print(data['TitleName']) Output: Corresponding paragraph......

[@karndeepsingh]

You can try to iterate both title name and Paragraph at the same time to generate respective title with paragraph. Haven’t tried it though. But it should be achievable.

[@DANstudiosable]

@@karndeepsingh Nope. The output titles and paragraphs are still random.

[@GuruTechHub]

hi. please make video on extract hindi table contains text in devnagri or utf-8 mangal to csv from images. i try lot on inter but not found any video or method.. please make video on this it will help lot

[@manuthvann7560]

hi Karndeep , I am totally appreciated your hard work a lot dear, but as I have followed up with your detectron2 vdo , there is an error with the Cuda of collab and PyTorch, I have tried to fix it but it doesn't work, can you help me out with that, the Cuda in my collab is 11.1 while in PyTorch is 1.7 , and I am working with GPU runtime. Looking forward to be hearing from you thx

[@karndeepsingh]

You have to pip install detectron 2 of pytorch 1.7 version similar to your colab version. You can change the pytorch version in the line where pip install detectron 2 is mentioned. Please check.

[@poojabhandari631]

how to do it in vs code

[@karndeepsingh]

Same as shown in video ☺️

[@mohitsoni1294]

can i save this extracted text to database?

[@karndeepsingh]

Yes you can extract and save it in database it required format.

[@HariHaran-mb3hh]

What if we don't know the language of the image content ?

[@karndeepsingh]

CV models never see language to detect the regions

[@HariHaran-mb3hh]

@@karndeepsingh yes..but after OCR process u can't copy the text. Its need to give language as a input for tesseract and ocrmypdf libraries. Then only u can copy the content

[@karndeepsingh]

@@HariHaran-mb3hh if you want it to detect the language before OCR then you can train an Object detection model and mark the region of interest with class as language (whatever text language present inside the region of interest) hence while inference you can get region of interest plus class as language.

[@A.El-Taher]

the sound of mouse click is very noisy