In order to enhance the efficacy of inspectors and disintermediate humans, the solution had to largely disintermediate humans and prompt human involvement on an exception basis only. The volume of documents being processed manually was a key indicator of a workflow that could be heavily automated. Moreover, detailed analysis through stakeholder interviews, research, and process examination indicated that the reduction of time spent with the various document types would free employees to focus on challenges less suited (i.e. less predictable and rote) to artificial intelligence. A comprehensive solution must address the problem holistically, necessitating the application of various forms of AI as well as a digitalization of process.
The promise of artificial intelligence is in its capacity for human enhancement, which differs significantly from the alarmist view perpetuated by the media. Artificial intelligence in its best application eliminates aspects of a human worker’s process that is better handled by a machine, rote and mechanical processes, for instance, while giving the human the bandwidth to pursue the type of work to which humans are best suited. It is rarely the case that human replacement is the end goal in the application of artificial intelligence.
To address a problem of this complexity, Hypergiant needed to leverage multiple techniques in order to be successful. Computer vision is used to identify and isolate individual documents within the order group. Leveraging optical character recognition (OCR) and fuzzy logic algorithms, the solution is able to detect and separate documents. The coupling of these techniques allows for automated processing of different sized order groups. In terms of automated field detection and reading, Hypergiant paired computer vision and OCR to determine whether key data in dynamically identified fields are readable as text – Green-colored fields are successfully read by the OCR system, and the yellow-colored fields indicate a low confidence. Prior to this process, a series of image processing algorithms are run to adjust and correct for watermarks, handwriting and scan warping. To minimize required human intervention, the solution utilizes field content validation, in this case by applying Natural language processing (NLP) & comparative analysis to translate the discovered and legible fields into computerized text that is compared between documents. Retrieving, reading, translating and validating documents are all done automatically and almost instantaneously. The user is shown the results of the document triage, alerting the user to action. The ability to robustly locate keywords for information extraction and the utilization of image segmentation for detailed items in Straight Bill of Lading documents was critical to success, as was addressing the challenges of isolating handwritten sections. Individual characters are segmented via contour matching and the convolutional neural network driving the computer vision capabilities addressed this challenge elegantly and effectively.