Image Processing: a Review on Industrial Applications

Affek, Michał and Barusco, Manuel and Dupļevska, Diāna and Freivalds, Kārlis and Hollósi, Gergely and Kofler, Corinna and Kopczyński, Michał and Almahasneh, Ruba and Newton, Philip and Strazds, Guntis V. and Susto, Gian Antonio and Szűcs, Gábor and Tatara, Marek S. (2026) Image Processing: a Review on Industrial Applications. INFOCOMMUNICATIONS JOURNAL, 18 (SI). pp. 25-43. ISSN 2061-2079

Preview

Text AIMS_SpecIss_ICJ_2026_1_2.pdf - Published Version Download (4MB) | Preview

Abstract

Image processing has become a key technology in industrial automation, supporting quality control, robotic manipulation, and process monitoring. However, deploying deep learning-based vision systems on the factory floor remains challenging. Industrial datasets are often small, imbalanced, and expensive to annotate, while production environments differ substantially from the everyday imagery used to pretrain most vision models. This paper reviews recent advances in industrial image processing and summarizes practical experience from several use cases developed within the AIMS5.0 project. We examine methods for synthetic data generation, multi-camera tracking, and continual learning, with particular attention to their role in data-scarce and evolving production settings. The presented applications cover robotic manipulation, quality inspection, assembly monitoring, and data annotation. For robotic grasping, a hybrid pose estimation pipeline combines AI-based detection and segmentation with Iterative Closest Point fitting. It achieves accuracy comparable to state-of-the-art approaches while requiring only an object point cloud instead of a textured CAD model and offers substantially faster execution. A segmentationbased method for bottle socket detection also provides promising localization accuracy, although further improvements are needed to meet strict mechanical tolerances. For industrial inspection, we show that class balancing, moderate augmentation, and transfer learning improve semiconductor defect classification under severe class imbalance. We further introduce a CycleGAN-based domain adaptation approach that is effective when source and target domains differ in their label distributions. In aluminium corrosion detection, supervised segmentation clearly outperforms unsupervised anomaly detection. Experiments on luminaire assembly show that video action recognition performs well on familiar products but generalizes poorly to visually different variants, illustrating the limitations of vision-only approaches in high-mix, low-volume production. Finally, an iterative pseudolabeling pipeline substantially improves object detection from a small initial set of annotated images. Based on these results, we identify several open challenges in this area, including data requirements, extreme class imbalance, reliable pseudolabel confidence estimation, automatic image-quality filtering, and scalable human-in-the-loop annotation.

Actions (login required)

View Item