Publication: Detection of label printing defects on sdd drive using pre-trained convolutional neural networks
Loading...
Date
2023-07
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Label printing defect is known as a cosmetic defect. Although this defect will not cause functionality failure of the Solid-State Drive (SSD) it will ruin the customer expectation of the product and the company's reputation. This defect can be identified through manual inspection, semi-visual inspection, and automated inspection. Manual inspection, in most cases, may lead to error detection and slow detection speed due to human error factors. Thus, many industries utilize semi-visual inspection and automated visual inspection to reduce both problems by applying the vision system. Knowledge such as machine vision and deep learning technology with the object detection technique is important nowadays to implement in the vision system for defect detection in the industry. In this study, the pre-trained convolutional neural network (CNN) with You Only Look Once (YOLO) network will be utilized to detect label printing defects using images provided by Western Digital. First and foremost, the region of interest (ROI) of the label printing will be extracted using the machine vision algorithm. After that, the annotation of defects and no-defects on the label printing is done using Roboflow. Then, the label printing dataset will be saved in the YOLO format and ready to be trained on Google Colab. Several versions of YOLO networks such as YOLOv5, YOLOv6, YOLOv7, and YOLOv8 are trained, evaluated, and compared. Then, the YOLO algorithm with the best performance will undergo hyperparameter tuning to further improvement. Finally, the YOLOv8 with optimizer of AdamW, the learning rate of 0.0001, batch size of 4, and activation function of SiLU yield the best overall performance with a precision of 0.978, recall of 0.906, F1-score of 0.941, mAP50 of 0.968, and mAP50:95 of 0.82. By having this high performance, YOLOv8 can aid the visual inspection in the industry with reliable performance.