Detecting surface anomalies with deep learningRački, Domen (Avtor) Skočaj, Danijel (Mentor) Tomaževič, Dejan (Komentor) Surface defect detectionvisual inspectionquality controlsolid oral dosage formspharmaceutical industrydeep learningsegmentationconvolutional neural networksvisual anomaly detectionDeep-learning-based approaches have proven to outperform other approaches in various computer vision tasks, making application-focused machine learning a promising area of research in automated visual inspection. In the dissertation, we apply deep learning to the challenging real-world problem domain of automated visual inspection of pharmaceutical products. We focus on investigating whether compact neural network architectures, adhering to performance, resource, and accuracy requirements, are suitable for application in the pharmaceutical visual inspection domain. We propose a compact and efficient convolutional neural network architecture design for the segmentation and scoring of surface defects, which we evaluate on a real-world dataset from the pharmaceutical product-inspection domain. In comparison with other related supervised segmentation approaches, we achieve state-of-the-art performance in terms of defect detection as well as real-time computational efficiency. When compared to the nearest best-performing architecture, we achieve state-of-the-art performance with a fraction of the parameter count, a multi-fold increase in inference speed, and increased surface defect detection performance. We furthermore propose an extension of the supervised approach to weakly-supervised learning via generated pseudo labels, which are obtained with an unsupervised deep learning model. The latter are utilized in a supervised fashion, since this allows the retention of compact model performance benefits, while removing or at the very least reducing the number of required manually annotated samples. Although this kind of approach is weaker when compared to supervised approaches using manually annotated labels, we achieve comparable results.20222022-12-05 15:38:43Doktorsko delo/naloga142942sl