In our work, we analyze how faulty weft threads in air-jet weaving machines can be detected using image processing methods. To this end, we design and construct a multi-camera array for automated acquisition of images of relevant machine areas. These images are subsequently fed into a multi-stage image processing pipeline that allows defect detection using a set of different preprocessing and classification methods. Classification is performed using both image descriptors combined with feature-based machine learning algorithms and deep learning techniques implementing fully convolutional neural networks. To analyze the capabilities of our solution, system performance is thoroughly evaluated under realistic production settings. We show that both approaches show excellent detection rates and that by utilizing semantic segmentation acquired from a fully convolutional network we are not only able to detect defects reliably but also classify defects into different subtypes, allowing more refined strategies for defect removal.
展开▼
