5
$\begingroup$

I am making an document parser which extracts data fields from the documents and store them in a structured way. Each field in my dataset is horizontal which is easy to extract.

enter image description here

But the model fails on following type of example -

enter image description here

Is there any way to extract invoice number and date from such images.

$\endgroup$
  • $\begingroup$ Does the document format remain constant? The positions of the fields? If then you can directly locate fields and then perform object detection to detect values from the images itself. $\endgroup$ – Sharan Jan 20 at 5:59
  • $\begingroup$ No it doesn't. Format changes as different companies have different type of invoices. $\endgroup$ – hR 312 Jan 20 at 6:00
  • $\begingroup$ what type of formats do you typically receive? Are they all text values, such as .doc and .pdf ? Or are there some elements that are exclusively image? $\endgroup$ – Leevo Jan 20 at 7:43
  • $\begingroup$ I receive images only like the images shared above. $\endgroup$ – hR 312 Jan 20 at 8:01
1
$\begingroup$

I think you already have some OCR in place? I don't know if you also have the x-y locations and size of the recognized texts?

I hope you have a model that knows (has learned) occurrences of 'invoice #' as a label.

And maybe you can machine learn to recognize values that could be invoice numbers. 2034, 200.00 could be invoice numbers, 'Date' and 'Service fee' not.

You could machine learn relations between objects, probably with the help of a distance function.

I would say that a string value that contains mostly digits, is near a label that matches 'invoice #', and also has a similar size, is the most likely invoice number.

564 could be an invoice number, but it is too far away from invoice # (further than 2034).

'Date' is close to invoice number, but it does not match an expected string for invoice numbers, since it is mostly letters.

$\endgroup$
  • $\begingroup$ Actually what I was planning to do is detect ` Invoice # 2034` as whole using some object detection technique and then detect the label and the value. won't that be an convenient approach $\endgroup$ – hR 312 Jan 17 at 7:47
  • $\begingroup$ And about your third point the invoice numbers can be of any format and can be any number I guess applying machine learning for the same isn't possible. $\endgroup$ – hR 312 Jan 17 at 9:18
  • $\begingroup$ @hR312, on the last point, it will be difficult for a machine to prefer '2034' over 'Date' for invoice number, since the transposed version of this mini table will be just as meaningful for humans.. $\endgroup$ – Pieter21 Jan 17 at 10:11
  • $\begingroup$ Then regex is the only solution that i can think of $\endgroup$ – hR 312 Jan 17 at 10:42
  • $\begingroup$ Will it be beneficial if i detect label and values from images iteself not from text $\endgroup$ – hR 312 Jan 17 at 10:53
0
$\begingroup$

A completely different answer:

I am currently following a course Computer Vision and Image Analysis: https://courses.edx.org/courses/course-v1:Microsoft+DEV290x+1T2020/course/

With your problem in mind you could follow along. Depending on previous knowledge you could skip a few sections. (I skipped immediately to Beyond Classification/Object Detection)

Globally, you could train an image classification model that could recognize regions of interest, with a classification of the content. Where the course addresses people, cars, buses, in your problem you have images, labels, content (of various types). You may need to experiment, and maybe have 'label-value-pair' as a class, or even 'label-multiline'. Or maybe labels and values separately work better? Or even a third option that you identify all combinations 'label' 'value' 'label-value'

You should probably define a low-wide boundary box for horizontal label/value pairs, and a more square one for vertical aligned label/value pairs.

You should end up with labeled regions of interest.

If you are happy with these regions, then the second step could be OCR. For the OCR you could use a similar problem division to recognize separate characters, and label the separate characters. And then you still have to combine the characters to words or values.

$\endgroup$
  • $\begingroup$ You should probably define a low-wide boundary box for horizontal label/value pairs, and a more square one for vertical aligned label/value pairs. I am trying this only but it requires a lot of data. $\endgroup$ – hR 312 Jan 20 at 13:21
  • $\begingroup$ @hR312, you can just take a few general rectangular shapes, not too similar. The image recognition algorithm will take care of scaling and shifting by nature. You will need to do a lot of tagging yourself or write programs that do the tagging for you. Given the complete picture and a few indicated spots of interest. $\endgroup$ – Pieter21 Jan 20 at 13:26
0
$\begingroup$

I would suggest that you should use a pre-trained OCR model and train your own custom model which only outputs required data.

Training method:

Just use a pre-trained OCR model like this, and remove the tail of the model and add your custom output layer with the required number of fields (in your it's case invoice and date). After this, freeze the head of the model and train your custom model with the data you have.

Note:

  • The accuracy of the model can be improved if you train this custom model by using as many different invoice templates as possible. Because this custom OCR model will have to learn to figure out the position of the invoice and date by itself.
  • If the model's output is incorrect for certain template you can always generate your own (synthetic) data by editing the template and add many examples of that particular template.

Using the pre-trained model, you can get pretty decent results with less training data. If you haven't used a pre-trained model, here is a more generalized way to use style transfer in PyTorch. I hope it will help you.

$\endgroup$
  • $\begingroup$ How will improving pre-trained OCR work for full problem? $\endgroup$ – hR 312 Jan 20 at 13:25
  • $\begingroup$ I didn't mention about improving pre-trained OCR model. I suggested using a pre-trained model to create a custom model that will do the required job i.e., extracting invoice no. and date. (the process of using a pre-trained model to create other models is called style transfer. It uses knowledge of one scenario and applies that knowledge in other application). $\endgroup$ – Vikas Bhandary Jan 20 at 13:38
  • $\begingroup$ A pre-trained OCR can tell what is in an image and localise where that text is. And we can use that knowledge to train a custom model to extract only invoice numbers shown in an image on a fixed location (similarly with dates or any custom data). $\endgroup$ – Vikas Bhandary Jan 20 at 13:41
  • $\begingroup$ If we train a model using this approach, then I suppose the model can learn that for each invoice template the invoice number is near the word "invoice". Whether it's placed horizontally or vertically, It wouldn't matter. $\endgroup$ – Vikas Bhandary Jan 20 at 13:56
  • $\begingroup$ Doesn't sounds feasible to me $\endgroup$ – hR 312 Jan 22 at 8:16
0
$\begingroup$

I have a similar use-case and a working product based on tensorflow object-detection api and pytesseract for OCR. On top of the extracted text, I perform regex for validation of the extracted information and cleaning it to meet requirements of other processes.

Steps:
1. Annotate images with some tool like labelimg.
I annotated a set of 1K images, similar to yours, with 23 different classes. The dataset is unbalanced with some classes appearing almost in every image to some classes appearing in only as few as 60. However, there are ways to ensure that this imbalance does not affect the performance of the network.
2. Choose a model from tf model zoo (I use this frcnn model) and retrain the last two layers using transfer learning.
3. Export the inference graph, perform object detection to identify the region of interest, and run OCR on the region of interest to extract the text.
I'd recommend storing the extracted data in a dictionary with class of the object as key and the extracted text as value.
4. Finally, have regex validate the text in the extracted field and perform any manipulation/transformation that is necessary.

The trained model can be deployed to production with help of tfserving. The same trained network can be deployed into a mobile app as well - look for tutorials on tensorflowlite for this.

Hope my answer helps you! I had a tough (but interesting) time gathering the knowledge required to get a production grade product that currently serves hundreds of request everyday. I would recommend reading completely all the links I have shared in this answer, and feel free for more questions. Good luck!

$\endgroup$
  • $\begingroup$ For IP reasons, I cannot share any code that was used for training or used for validation, etc. However, they are all out there and most of them are covered in the shared links. $\endgroup$ – AccLok Jan 22 at 23:04
  • $\begingroup$ what other then regex can we use? can we train a line detector that gives both vertical and horizontal label value as a line. what I mean is Invoice # 2034 in image below will be detected as a line. $\endgroup$ – hR 312 Jan 23 at 12:02
  • $\begingroup$ Annotate the portion that has Invoice # XXXX as a separate class like 'invoice', Order # / XXXX as 'order', etc. With enough training samples, the model will be able to identify different regions as different objects. While classification predicts the class, regression gives the xmax, ymax, xmin and ymin coordinates of the detected class. This can in turn be fed to to the pytesseract for extraction of text only from the area of interest within the coordinates. $\endgroup$ – AccLok Jan 23 at 14:16
  • $\begingroup$ If you found any part of my answer to be resourceful, please upvote the answer. $\endgroup$ – AccLok Jan 23 at 14:16
0
$\begingroup$

More or Less this would be helpful

Link: Extracting information from documents

Approach & Algorithm from the above blog

Approach

The algorithm looks for phrases that look like a date. Then it picks the one which appears in the highest position in the document. In the corpus we used, almost every date contained the month written as a word (e.g. April), the day written in digits (13) followed by the year (1994). Sometimes, the day was printed before the month (e.g. 4th September, 1984). The algorithm looks for the patterns M D Y and D M Y where M is a month given as a word, D is a number representing the day and Y a number representing a year.

Software Tools

Our implementation runs in a Jupyter Notebook with Python 3. We use Tesseract version 4, for doing OCR through the wrapper pytesseract. Since the software sometimes gets a letter of the month wrong (e.g., duly instead of July), we accept all strings which almost look like a month in the sense that only a few letters need to be changed to reach a valid month. The number of these operations is called the Levenshtein distance, a common string metric in natural language processing (NLP). For example, the Levenshtein distance of duly and July is 1. Similarly for Moy, Septenber or similar errors. We use python-Levenshtein. For detecting numbers (years and days), we use regular expressions. We process all the tables in Pandas and use tqdm to have a neat progress bar.

Algorithm step by step

Similar Questions from Stackoverflow:

  1. https://ai.stackexchange.com/questions/16076/how-to-extract-information-from-the-image
  2. How to Extract Information from the Image(PNG)
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.