Tagging or marking goods is essential for the warehouses, quality assurance, or the automatization of industrial processes. Serial numbers are embedded into RFID tags, encoded into a QR Code, and usually processed by a camera or laser scanner. There are industries, however, where these traditional methods donot satisfy the requirements and alternatives are desired. This paper focuses on the feasibility of using deep learning technology to read serial numbers from a steel bar which was printed with English characters by a needle printer.
The detection and recognition of text is a well-studied computer vision field also known as optical character recognition (OCR). In this work, we demonstrate that existing OCR methods are unable to solve the posed task without additional training of the deep learning models. This work divides the problem into three individual sub-problems and approaches all three of them by using deep learning technologies. The given dataset was analyzed and divided into training and validation sets for each individual problem, while a part of the entire dataset was reserved for the final system evaluation. After selecting the best model for each subproblem, the resulting system could achieve a serial number accuracy of 90.4 percent and a false positive rate of zero percent. This work has shown that deep learning technologies can be used to read serial numbers, but it is essential to include a checksum to be able to verify a prediction.
Det finns övrigt digitalt material (t.ex. film-, bild- eller ljudfiler) eller modeller/artefakter tillhörande examensarbetet som ska skickas till arkivet.