Скачать или смотреть Image Analytics : The technical flow of business use case (Aritha Webinar)

In this webinar by Aritha, Mr Pruthviraj Phanendrakumar, a data scientist and tech enthusiast, detailed on a many aspects of application of Image Analytics in general and the OCR tools and techniques in particular. Image analytics attempts to replicate the possibilities of human eyes with the computer. It’s application could be as simple as an OCR or as complex as analysing micro-seismic images which helps us predict natural disasters like earthquakes. In this webinar, a use case of developing an AI application used for extracting key fields from scanned images was explained.

Pruthvi drew from his personal experience of working with a large organisation where the document scanning and update was happening manually (keying data into excel and updating database) for millions of records. That is when the data science team was consulted for building a homegrown AI based application for automating the process. Text extraction was achieved using the open source library “pytesseract” in this project. Text Extraction Pipeline with text detection done using Deep learning technique was explained in detail. The text extraction pipeline contains 5 steps:
1. Labelling required fields from the document
2. Training the model with various formats of the document
3. Detecting the fields
4. Applying OCR on detected fields
5. Writing to CSV
In this use case, the detection accuracy was 90 - 95% as deep learning technique was used. There are expensive tools to achieve extraction, like the Amazon’s “textract” tool, for more accurate results and more complex requirements like extracting handwritten text.

Three distinct advantages of implementing such a solution are:
1. It saves time and has monetary benefits
2. Provides Higher accuracy than manual approach
3. Serves as a primary input for downstream applications

Pruthvi walked the audience through various other tools available for OCR like the OCROpus, Ocurlar and SwiftOCR apart fromTesseract. He concluded the talk by listing some of the limitations of using Tesseract. The session was lively with participating audience.

Some interesting questions answered during the session were:

Will it be able to extract text from scanned PDF of handwritten documents?
Are there any preprocessing techniques used in OCR?
How much time does it take to process from end to end? Is it done as batch process? Is it done realtime?
Is there any mechanism to retrain the model periodically using automated ML Ops techniques?
Is watermark recognition and removal possible with this tool?