Investigation of Augmentation Impact of Fibers Macro-Images Set on the Neural Network Classification Accuracy
Вантажиться...
Дата
2026
Назва журналу
Номер ISSN
Назва тому
Видавець
Анотація
The paper investigates the impact of image augmentation on the accuracy of neural network classification of textile fibers macro-images. The study focuses on improving automated recognition of fiber composition for textile sorting and recycling systems using computer vision and convolutional neural networks.
Several augmentation techniques, including rotation, zoom, translation, contrast adjustment, and horizontal flipping, were experimentally analyzed using a MobileNetV2-based classifier trained on a three-class dataset of textile fiber macro-images. The results demonstrated that augmentation can improve classification stability and generalization, although its effect differs between classes and transformations.
The study showed that zoom and translation provided the best overall improvement in macro-F1 score, while some transformations negatively affected specific classes. The obtained results confirm that augmentation in texture-based textile classification should be treated as a controlled experimental factor rather than a universal preprocessing operation.
Опис
Ключові слова
Бібліографічний опис
Savenko B., Zalutska O., Mazurets O., Molchanova M. Investigation of Augmentation Impact of Fibers Macro-Images Set on the Neural Network Classification Accuracy. Proceedings of VII International Scientific and Practical Conference «Research in Science, Technology and Economics». May 6-8, 2026. Luxembourg, Luxembourg. Pp. 255-261