Glove defect detection via YOLO V5

• Main Authors: Yong, Chen How, Ahmad Fakhri, Ab. Nasir, Khairul Fikri, Muhammad, Anwar P. P., Abdul Majeed, Mohd Azraai, Mohd Razman, Muhammad Aizzat, Zakaria
• Format: Article
• Language: English
• Published: Penerbit UMP 2021
• Subjects: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://umpir.ump.edu.my/id/eprint/33977/1/Glove%20defect%20detection%20via%20YOLO%20V5.pdf; http://umpir.ump.edu.my/id/eprint/33977/; https://doi.org/10.15282/mekatronika.v3i2.7342; https://doi.org/10.15282/mekatronika.v3i2.7342

Malaysia is one of the biggest producers and exporters of gloves in the world. To meet and exceed the customer’s expectation, a predictive defect model is necessary to minimize the defect glove. There are three crucial parts to develop an effective defect glove detection model, which are data collection, model development and model evaluation. The data provided should be good quality, the algorithm for developing the model should reach high accuracy and high inference time due to the fast glove production line, and the developed model must compare to the other quality model to prove its robustness and effectiveness. This paper focuses on employing the YOLO V5 model for glove defect detection as well as investigating the efficiency of other several deep learning approaches. The dataset collected in this research was 493 images with three classes which are normal glove, tear glove and unstripped glove. To avoid overfitting due to the small amount of dataset, argumentation processes such as saturation, exposure and noise were applied to increase the dataset number to 1148 images. Data were then split to 70:20:10 for the training-validation-test ratio. The parameter setup was 100 epochs with 129 iterations. The YOLO V5 was compared with Scaled YOLO V4, Detectron2 and EfficientDet by the training time, model size, accuracy, and inference time. In conclusion, the best model was YOLO V5 because it reached the lowest training (0.259 hour) and inference time (0.0095 seconds), smallest model size (14418kb) and highest accuracy (mAP = 0.9951).