Delays in translating information from old documents and technical drawings can hamper the distribution of data. Epitomy Solutions is now integrating existing optical character recognition (OCR) technology with the latest artificial intelligence (AI) solutions to give organisations more efficient access to rich product data.

For many companies, large amounts of valuable data still exists on printed sheets and old schematics. Although scans allow these assets to be included within online systems, the information is often hard to read and does not allow click-through or easy identification.

Getting good quality images that can be indexed is the main objective. Achieving better first-time recognition of characters will improve a company’s ability to digitise all information held on its systems, and to process that data quickly.

Once equipped with otherwise obscure and unusable part references or product descriptions, a company’s data management system can then deliver key information to front end tools being accessed by customers.

Epitomy’s in-house Image Processing App has successfully achieved this goal in most cases, but OCR has always had its limitations.

Existing OCR route

OCR is the conversion of images and characters that have been typed, handwritten or printed into machine-encoded text. This is done by referencing against learned character sets. The quality of files (often in PDF format) has always dictated the success rate of recognition. Manual intervention is often necessary to generate the most accurate results.

Digitising printed texts allows system users to electronically edit, search and display on-line materials. The resultant data can be used in machine processes such as cognitive computing, text-to-speech and text mining.

Where conversions are successful, software exports coordinates of the recognised text. This reference point can then be linked to other data sources, whether that is a product image, description, technical drawing, or pricing and availability information held within the central database.

Where image quality was particularly poor, often on old printed documents that had been poorly scanned, an accurate interpretation might not be possible, potentially leaving a block of product data missing.

This void of information could have implications for organisations seeking to use or sell on older products that are still required on machinery or vehicles. Companies have been forced to create time-consuming workarounds to bridge the gaps.

The future of static data interpretation

A desire to deliver enhanced results in the most challenging circumstances has encouraged Epitomy Solutions to investigate artificial intelligence solutions in conjunction with existing products. Research by a number of technology companies has begun to challenge the notion that machines cannot improve what the human eye and associated brain processes already offers. The tables have turned, and learning algorithms are starting to put computers in the driving seat.

To complement the existing OCR offering, Epitomy Solutions has explored what Google’s TensorFlow™ could provide our clients. It’s an open source software library for numerical computation using data flow graphs. Crucially, like all evolving AI technologies, the system is always learning as it consumes fresh data. TensorFlow™ is flexible and can be used to express a large variety of algorithms; this includes training and inference algorithms for deep neural network models.

TensorFlowMnistGraph
High level graph of a TensorFlow™ Program for identifying handwritten numbers

For the team at Epitomy, TensorFlow™ is downloaded into the Python library before coding consolidates the technology alongside our in-house App that drives the existing OCR code. Our clients will still use the App we developed to enable them to scan and digitise documents, but the TensorFlow™ code kicks in if recognition of characters cannot be completed by the original solution. It’s the best of both worlds; old and very new sitting side by side.

The deep neural network returns a classification that gives a best fit. Effectively, to quote an example, the results tell us that the loss function (that is, the chance of the scanned result being wrong) is a given value, say 0.03 (or 3%). Turned around, we discover that there is a 97% chance that the characters in question have been successfully identified. That’s typical of most scans, and better than the average human eye.

Getting the data customers need

There are many potential applications for such artificial intelligence. For Epitomy customers, where manufacturers often list thousands of parts, a field operative could harness the technology via a mobile phone or tablet. By taking a photo of a broken or failed part, this data can be sent back to the central Product Information Management (PIM) system. Photos will be matched, part numbers will be cross referenced and known errors assessed. Data will flow across platforms.

The user will then receive accurate part information, price, availability and even a timescale for delivery and fitting. For older parts, scanned images will have linked to more expansive information. There could be workarounds, or alternative products could be suggested from linked catalogues.

Merging technologies will help end users and customers get the product information they need to make informed decisions. That’s the benefit of fully integrated data management systems. Ensuring that accurate information is fed into the system, the right data can be delivered back to users on multiple devices.

Epitomy Solutions is always looking to harness the power of new technologies to provide the most efficient and productive data management systems and front end interfaces to make the lives of users and customers easier.