Understanding IDP Capture Preprocessing

Understanding IDP: Capture and Preprocessing

by Sujith Parakkunnath, on November 11, 2021 9:00:00 AM PST

Gartner recently released, “Infographic: Understand Intelligent Document Processing”. According to Gartner, “The market for document capture, extraction and processing is highly fragmented. Data and analytics leaders should use this research to understand the process flow and differentiated capabilities offered by intelligent document processing solutions.” In this series of posts, we will speak to the 6 critical flows in IDP that Gartner covers and how Infrrd solutions stack up.

1. Capture or Ingestion
2. Document Preprocessing
3. Document Classification
4. Data Extraction
5. Validation and Feedback Loop
6. Integration

In this first post, we explore Capture or Ingestion, as well as Document Preprocessing.

If you are in the business of document processing where a large volume of documents are processed - be it invoices, receipts, insurance forms, or healthcare documents - you may be considering migrating or have already migrated to an automated process to remain competitive. This is where Intelligent Document Processing, or IDP, solutions can be a powerful way to transform your business.

IDP is the buzzword when it comes to document processing automation. With IDP solutions, documents are processed at an exponential pace and high level of accuracy aiming at touchless strategy, which means minimal manual efforts.

IDP focuses primarily on gaining meaningful inferences and relevant information from variations in documents from different document types, such as unstructured, semi-structured, and structured.

Preprocessing in IDP solutions can be broadly categorized into two areas

1.  Automated capture or ingestion
2.  Automated preprocessing

Document capture or ingestion

The first phase of document processing by an IDP solution is smart capture or ingestion. OCR and powerful machine learning algorithms play a major role here.

In simple terms, your business is an accumulation of data, starting from a client email to complex inventory or invoice documents. You can broadly classify this data into three categories:

  Structured: This is the most organized data where information is well structured. For example, a government form or tabular data in an Excel sheet.

  Unstructured: This is unorganized data and is the most complex to process. Information may be in different formats, such as an image, a video, an audio file, an email message, or even news articles.

  Semi-structured: This is a blend of structured and unstructured data. For example, invoices, purchase orders, bills of lading, contract documents, company annual reports, and so on.

In the real world, more than 75% of data is either in an unstructured or semi-structured format. To automatically process and structure this data competitively for a large volume of documents, you need a robust and efficient IDP solution, such as Infrrd’s IDP, so that you do not compromise on time, accuracy, and efficiency as compared to your competitors.

So, how does an AI-based IDP solution handle data, especially unstructured and semi-structured?

IDP systems use AI-based optical character recognition (OCR) which is used with technologies, such as computer vision and natural language processing(NLP). OCR is primarily used to detect language-related characters, letters, numbers, etc. Traditional OCR has a high dependency on structured data. However, with computer vision and NLP, the capabilities of OCR to handle unstructured and semi-structured data have undergone a paradigm shift.

IDP solutions recognize the data extracted by OCR and enable you to structure it based on the data types, irrespective of whether the documents are structured, unstructured, or semi-structured. Moreover, machine learning algorithms in IDP systems enable the system to exponentially learn from training and corrections each time a document is processed.

Document preprocessing

In IDP solutions, documents pass through the preprocessing stage where OCR and machine learning algorithms evaluate the document quality and implement resolution measures before actual extraction. Document preprocessing is an important step to ensure high-quality extraction. It refers to cleaning, organizing, and transforming the raw data to match the quality expected or mandated by the IDP or machine learning models. It is primarily a data mining technique aimed at improving document quality.

Infrrd’s IDP solution uses advanced AI technologies for document preprocessing. Let’s dive into the preprocessing strategies of IDP solutions.

Data annotation and labeling

Data annotation and labeling is a configuration process where you are setting up your system to make it preprocessing-ready.

IDP systems have a machine learning-first approach. This means that the documents are processed based on a configured machine learning model. Infrrd uses training, also known as tagging, to configure models for document processing. The configuration user or manager is equipped with data and annotation capabilities to train a specific document type, including the document fields. The data annotation or tagging is minimal when you use a pre-trained global model which has the potential to process documents with high accuracy for known document types. However, the data annotation features of Infrrd’s IDP enable you to easily configure, tag, and train even unknown document types.

Merge or split documents

When you are processing hundreds or thousands of paper or digital documents, it is the practical reality to have numerous unstructured, unorganized multipage documents for processing. This is where the key preprocessing activities are automatically performed by IDP platforms, one of the keys is the merging or splitting of your documents.

Infrrd’s IDP solution uses technologies such as computer vision and NLP to analyze your documents and recognize their structure and layout to detect the documents that need to be split. You can upload a parent file and Infrrd’s system will recognize them, and then split them into individual documents to enable the inbuilt OCR engine for downstream processing, as shown.

Infrrd OCR Engine Dashboard

Skew correction

IDP systems offer the best extraction results when document elements, such as images and text, are in an upright position. However, in the real world, this scenario is far from true as you have multiple variations of documents, including unstructured physical copies, where many of the elements appear skewed. The text appears rotated or tilted in different angles. Think scanned documents. You need to correct the angular tilts and orientation before processing the document - literally to zero degrees. One of the popular techniques used by IDP solutions is skew correction, also known as deskewing. This process eliminates the manual efforts as skew correction is automated in IDP.

Infrrd’s IDP solution uses a combination of coherent machine learning algorithms, deep learning models, OCR capabilities, and deskew techniques and mechanisms, such as Hough Transform, to ensure exceptional skew correction results. The high-level flow of our deskew techniques include:

  Identifying the text blocks or images to be skewed
  Detecting the skewness and calculating the text angle
  Applying Infrrd deskew mechanisms to correct the skew to zero angles

Deskew mechanisms to correct the skew to zero angles


Denoising is the process of detecting and removing the grainy sections, such as black dots, blur, or shadows, and cleaning them up to restore better quality. The physical documents you upload may have deteriorated over time because of any number of reasons, such as dirt, stains, or wrinkles. This is where denoising becomes an important preprocessing activity for IDP systems.

Some of the IDP methods to denoise documents include:

  Median filtering: Primarily to peel out the white spaces in the background

  Edge detection, dilation, and erosion, also known as the EDE method: Detects unwanted elements, such as boxes or lines, and removes them 

  Adaptive thresholding: Identifies unwanted elements by detecting the threshold value of pixels 

  Linear regression: Performs a brightness and contrast correction by predicting the pixel intensity and linear relationship between sections that are clean and dirty  

  Auto encoding: Segments data into layers to remove irrelevant elements

Infrrd’s IDP uses multiple techniques to denoise documents, some of which include adaptive thresholding, binarization, and Gaussian smoothing, also known as Gaussian blur or Gaussian function. Our unique and constantly evolving skew correction strategy focuses on transforming low-quality documents to comply with Infrrd’s extraction-readiness state.

Denoising the Document

Data validation and correction

Data validation is a process to check the compatibility of the document, such as whether the document is in a valid format or if the resolution matches the defined specifications. Moreover, these are activities to define a certain level of quality for the raw data, such as identifying discrepancies between training set data and the data you need to process. The system automatically corrects most of the discrepancies identified at this stage. For example, if a document or an image you upload was captured in an old phone with low lighting, it will probably have low resolution, maybe 100 dots per inch (dpi). The Infrrd IDP system will automatically detect issues in such documents and increase the resolution before processing it, perhaps to 300 dpi.

Taxonomies, ontologies, tags, and more…

Another aspect to consider in an IDP solution are the options to integrate with taxonomies, ontologies, or tags. Taxonomies refer to the bottom-up approach of NER models to collect data in categories. Machine-language ontologies refer to identifying the pattern of various entities and relationships among the data. Tags refer to a key-value pair approach for data set tagging. The right set of tagging enables an AI system to be more effective and improve accuracy.

When all is said and done, make sure the IDP solution you invest in has exemplary capabilities for document preprocessing like Infrrd’s IDP. We recommend you make this a checkpoint when you evaluate or choose your IDP technology partners.

In our next post, we explore Gartner’s description of Document Classification, and how Infrrd stacks up.

Source : Gartner, Infographic: Understand Intelligent Document Processing, Shubhangi Vashisth et al., 22 September 2021

Topics:Intelligent AutomationAI ReadinessBusiness InsightsIntelligent Document ProcessingOCR Alternative

About this blog

AI can be a game-changer, but only if you know how to play the game. This blog is a practical guide to turning AI into real business value. Learn how to:

  • Make sense of complex documents and images.
  • Extract the data you need to drive intelligent process automation.
  • Apply AI to gain insights and knowledge from your business documents.

Explore Our IDP Platform

Subscribe to Updates