PUB - Publikationen an der Universität Bielefeld

The detection of defects is an import part of quality management in many industrial processes, since defective products can either be repaired or sorted out at an early stage. One application area is the detection of stains and holes on laundry after washing in order to sort out items of laundry containing holes or to rewash them if they contain tenacious stains. Little research has been conducted in this field to date because laundry is crumpled during and after washing so that it is impossible to capture images where defects are visible. However, there are research efforts, such as the Flachwaschen Project, that look at washing laundry flat and thus allow its automatic inspection. In this context, the doctoral thesis at hand investigates the use of transfer learning with deep learning methods for salient object detection to detect defects on patterned laundry. To this end, a custom dataset of images of laundry was created, and images were labeled pixel by pixel as defective or clean. Based on this dataset, it is shown that salient object detection is a good fit for this task, as defects are detected and patterns are ignored. It becomes apparent that the common measures for salient object detection are unsuitable for an application-oriented evaluation. This is why a custom measure is developed. Using this measure, it is ascertained that the addition of a backlight to the image acquisition enables the detection of defects on the back of laundry. Furthermore, it is shown that salient object detection methods generalize well to unknown patterns, and if there are structures that cause errors, a quick fine-tuning with little data is sufficient to improve performance. Moreover, results reveal that pixel-wise labeling by humans is error-prone, which means that defects are labeled roughly or are overlooked entirely. It turns out that the inaccurate labels are the main bottleneck for the defect detection performance and the interpretability of the evaluation. Consequently, it is shown that a self-training approach can be used to improve the labels by adding overlooked defects from a model’s predictions, which in turn boosts model performance.