A DEEP LEARNING SEGMENTATION APPROACH FOR AUTOMATIC INVOICE IDENTIFICATION

Abstract

The research on object detection has come a long way and has in recent years become so efficient andreliable that we see many practical applications in simple detection of objects in pictures to advancesegmentation in video recordings. Despite its popularity, there are no de facto approach availableand several methods are competing to be superior.This thesis focuses on object detection in photos where the desired object is letters and numbers, andthe photo is a PDF-file of an invoice without any use of OCR and any knowledge of the layout. Theresearch compares Tensorflow’s Object Detection API with SSD Inception V2 and YOLO. One of themethods shows promise, though it will need some more work. This thesis shows that SSD seems toneed more training, possibly with more varied data, while YOLO seems to not be a good method forthis type of object detection.The results from SSD varies as it sometimes get the correct location, though not always being certainin its decision. For other invoices it misses completely, and might guess at other short words ornumbers. Sometimes the words and numbers make sense, and could be the start of understandingwhat it is looking for. The tests done with YOLO gives results where it guesses almost the samelocation for every invoice, regardless of where the object actually is. It seems to overfit on the mostpopular formats, which is going against what this thesis is trying to do.