The Effect of Data Augmentation on Accuracy Values In Fabric Defect Detection

Objective: The purpose of this study is to use computer vision to replace the role of operators. The implementation is carried out by integrating the inspection tool prototype with computer vision. Theoretical Framework: The textile industry is highly dependent on fabric quality, because it has an impact on the selling price which can drop 45-65% from the normal price. Fabric inspection is very important in controlling the quality of fabric before being marketed. Currently, fabric inspection is carried out manually by the operators so that it is less effective in detecting fabric defects. Method: This study employs the YOLOV8 architecture with a dataset of 264 training images, 75 validation images, and 38 testing images, comparing three different augmentation: (1) flip and exposure, (2) flip, exposure, and blur, and (3) flip, exposure, blur, and mosaic on the Roboflow platform, followed by model training to detect fabric defects. Results and Discussion: The results show the best augmentation in the first augmentation training with the highest precision value of 73% and the lowest class loss value of 0.49. The model achieved a mAP 50 value of 0.65734, which is the second highest, while the highest mAP 50 value was obtained in the third augmentation training, reaching 0.80128. Research Implications: This study demonstrates that applying computer vision with YOLOV8 can improve fabric defect detection compared to manual inspection. The use of proper data augmentation techniques significantly affects model performance, where simpler augmentations yielded better precision in some cases. Originality/Value: This research contributes to the textile industry in INDORAMA by offering an automated fabric defect detection method using computer vision and deep learning. The findings provide insights into the effectiveness of different data augmentation techniques in training object detection models for fabric inspection.