Abrage Defect Detection Using Transfer Learning Methods

This study presents a deep learning-based approach for the automatic detection of Abrage defects in the textile industry, which arise due to variations in dye uptake or irregularities in raw materials during yarn production. A review of the literature reveals that while there are numerous studies on weaving or knitting defects on textile surfaces, there is a notable lack of research on the detection of Abrage defects—which occur on yarn bobbins and directly affect product quality. The aim is to develop a system that provides objectivity and continuity, replacing traditional control processes that rely on human observation and are costly and prone to high error rates. Within the scope of the research, a unique dataset consisting of yarn bobbin images captured under UV light in a controlled experimental environment was utilized. Preprocessing steps such as grayscale conversion, background removal, and masking of conical sections were applied to the images. A dataset consisting of 235 images was created using data augmentation methods. In the study, transfer learning models—including InceptionV3, DenseNet121, VGG16, MobileNetV2, Xception, EfficientNetB3, and InceptionResNetV2, all pre-trained on ImageNet weights—were compared alongside a CNN architecture. Experimental results were analyzed using accuracy, precision, recall, and F1-score metrics. The Xception model demonstrated the highest performance with 91% accuracy and 93% recall, emerging as the most ideal solution in terms of speed-performance balance.