Sistema de inspeção de etiquetas por visão computacional e aprendizado profundo na indústria 4.0

Abstract

Product inspection is an essential step in manufacturing processes to ensure the quality of the final product. Traditionally, this inspection has been done manually by human operators, which is time-consuming, expensive, and can lead to errors due to human subjectivity and fatigue. In recent years, most visual processes in a factory are being replaced by computer vision techniques. With the advances in deep learning approaches, optical character recognition and object recognition are technologies that can be used in different scenarios. In this study, a methodology capable of extracting textual and non-textual information applied to modem labels is developed. The proposed method consists of the following components: two object detectors that perform label detection and simultaneous QR code and barcode detection, both using YOLOv5; a content decoder for QR code and barcode using Zbar; an OCR system using PaddleOCR; and a set of rules applied to post-processing of the information. To do this, three datasets were created: the first containing images of modem labels to train the label detection model, the second containing a mixture of label images and various environments containing QR code and barcode to generate the QR code and barcode detection model, and a ground-truth base containing modem label images with their expected system outputs. The proposed system was evaluated with different label models, and its execution was done on a CPU. The label readings achieved average values of 0.21% Character Error Rate, 2.16% Field Error Rate, Label Accuracy of 76.19%, and execution time of 2.79 seconds for the first model, and 0.04% Character Error Rate, 0.62% Field Error Rate, Label Accuracy of 92.50%, and execution time of 1.73 seconds. Experimental results show that the developed solution can be used in production with high accuracy rates and significantly better execution time than a human operator.