Hyperscalers’ PaaS platform versus SAP’s BTP (ex SCP) PaaS platform, which to choose ?
Recent years have seen the emergence of generalist Paas platforms such as AWS, Azure, Google and more recently the Paas platforms of major software publishers such as SAP.
In addition, there is an abundant supply of open source bricks.
This offers companies a very wide choice to meet their specific application needs, but it also complicates their choices and we note that, faced with pressure from suppliers seeking to promote their solutions, IT departments find it difficult to see clearly.
Through a particular service offered by these platforms: character recognition (OCR), the objective of this article is to highlight the main characteristics of 3 approaches and to show that the choice between them is ultimately a « make or buy » alternative.
It is common in cross-functional company processes for the original media to be in paper, image or PDF format (e.g. an order received from a customer, a supplier invoice or a receipt). However, the company’s management applications process structured data, and without structured data to feed them, this type of tool would be out of date. To process text data, it must first be extracted from its original medium before being transferred from one interface to another in the management IS. To solve this problem, there are text recognition tools, known as OCR.
These tools extract the required data in a format that can be used by business applications. To do this, they proceed in two steps :
- Recognition and extraction of characters from the source document
- Semantic analysis of the generated file to identify relevant fields
So which solution should you choose ?
We will determine here the characteristics of these different solutions, to enable you to determine which one would be best suited to your use.
Among the various OCR solutions, we can distinguish between open source solutions (to be hosted on a local server), OCR services offered by generalist PaaS platforms (Azur, AWS, Google, etc.), and OCR services offered by the PaaS platforms of software package publishers or publishers of specialised business solutions such as SAP.
1) RECOGNITION AND EXTRACTION OF CHARACTERS FROM THE SOURCE DOCUMENT
Open Source OCRs :
For this first step, they generate a text file, from a scan of the document. This is a solution that has the advantage of being free in itself, as there are open source libraries on python. It also does not require a connection to the cloud. However, it has random results in terms of quality. If the initial document is blurred or encrypted, the quality of the resulting text is very random.
Generalist Paas (OCR cloud raw data extraction)
This solution generates a .Json file, which is more accurate than the text file, making it a much more accurate tool than local OCR, with good performance even with blurry or badly scanned documents. It is a remote tool, requiring a connection to the cloud. The scanned document is sent to the cloud OCR, which responds with a .Json file. The advantage of this type of file is that it is more accurate than a text file, as it contains not only characters but also information about the position of the text on the page in the form of bounding boxes, which makes it possible to manipulate the coordinates. Some versions provide another useful piece of information: a character detection confidence index, which allows the probability of error to be judged. Information about the layout of elements on the page is hierarchically arranged at different levels of precision: paragraph, line, word, and character for example. There are several APIs to implement this solution, such as G vision and MS Azure. They both offer a good level of performance, with data extraction in less than 5 seconds. Other APIs such as Recognition by Amazon are less efficient.
Paas platforms of software publishers. ( Turn-key cloud OCR)
Turn-key cloud OCR does not require this first step. They take care of all the work in the chain and directly restore the information of the requested fields.
2) Semantic analysis of the generated file to identify relevant fields
- Open Source OCR :
In the case of the local solution, this requires the writing of a specific template file that allows to query the searched information. Additional work is required, but programs are available in open source under python.
- Generalist Paas platforms (OCR cloud extraction of raw data):
For the classic cloud solution, after generating the .Json file, a custom script must also be written based on the coordinates of the information to be searched in the document. However, the nature of the .Json file makes this information easier to extract than in the case of open source OCR.
- Paas platforms of software publishers. (OCR cloud turn-key)
The turn-key cloud solution directly returns the specific fields requested. Some APIs also offer artificial intelligence type learning with neural networks, which improves performance. This is the case with Nanonets and Verify. Other APIs, such as Document information Extraction by SAP, do not require any learning, and allow information to be extracted directly from fields with the same reliability as the generalist Paas platforms. However, in France, SAP’s solution is only suitable for certain invoice-type documents, even though it is expected to be open to other types of documents in the medium term.
This solution is very efficient, with results similar to those of classic Paas platforms, but it does have some drawbacks. Indeed, although it does not require any integration effort, the lack of access to the code has a disadvantage: in the event of a bug, it cannot be solved directly by the IT team, but only the editor is able to intervene.
3) Cost Structure
- Open Source OCR:
For this solution, for a lower performance, the only cost is fixed: a non-negligible integration cost in order to write the template file for retrieving the desired data. This solution is the least efficient, but as soon as a volume is reached for which its integration cost is lower than the invoicing of the software package publishers’ Paas platform, it is the least expensive of all.
- Generalist Paas platforms (cloud OCR for raw data extraction:
For this solution, there is a non-negligible fixed integration cost in order to develop the in-house script for retrieving the desired data, and a variable cost for invoicing the cloud service, depending on the number of documents processed (relatively accessible: around $1.50 per 1000 documents scanned).
Thus, when the volume of documents is significant, it is less expensive than the solution of the Paas platforms of software publishers, for an equivalent performance.
- Paas platforms of software publishers. (“Turn-key” cloud OCR)
For this solution, in the absence of any integration effort, the only cost is that of invoicing, depending on the number of documents processed.
Thus, for a low volume, the Pass Platform from software publishers is both the least expensive and the most efficient solution. On the other hand, for a high volume, it is the most expensive of all.
It should also be noted that if the expected use falls outside the field for which the platform was designed, the added value provided by the semantic analysis proposed by the publisher becomes irrelevant and the platform loses its interest compared to generalist Paas platforms. For example, such a platform can save a company a lot of time when recognising supplier invoices if it has been designed for this purpose, but will lose its interest if the company wants to use it to recognise cash register receipts if the platform has not been designed for this purpose.
What to choose ?
FOR A LOW VOLUME or fairly standard requirement
Low enough that the cost of the Paas solution from software publishers falls below the higher integration cost of the generalist Paas solution, or even the Open Source OCR solution.
FOR HIGH VOLUME or very specific needs
High enough that the cost of the software package Paas solution is higher than the integration cost of the general purpose Paas solution and the Open Source OCR solution.
In summary, the choice between an open source solution, a general-purpose Paas platform service or a Paas platform service from business software publishers can be summarised as a « make or buy ».
Open source solutions have a low entry cost but require a significant integration effort on the part of the company, and the maintenance and operation of the solution is completely at the company’s expense. In our opinion, this approach can be justified for very specific needs or large volumes that allow the investment costs to be amortised.
The Paas services of software publishers such as SAP, on the other hand, have a much higher cost per use but are turnkey solutions that require little integration effort on the part of the company as well as little maintenance and operating effort because these are taken care of by the service provider. These services are relevant when one or more of the following criteria are met: need to implement the solution quickly, moderate volumes, fairly standard use cases covered by the platform.
The PaaS services of generalist publishers fall between the two; they allow the company to be relieved of the raw recognition component for a fairly moderate cost per use. On the other hand, the company will have to invest in an integration effort to deal with the semantic analysis dimension (via developments or possibly by using additional services offered by the SaaS platforms). This approach is often a good compromise because it allows the company to benefit from the scale effect and robustness of the PaaS services of generalists and to concentrate on its specific component, which is generally semantic analysis.