Comparison between AI-based and traditional computer vision algorithms for pharmaceutical stoppers and caps sorting

Industrial vision systems are increasingly employed in automated production lines to replace repetitive human tasks with precise, reliable, and high-speed operations. This thesis presents the development of a PC-based vision system for the automatic inspection and counting of pharmaceutical stoppers and caps, designed and implemented in collaboration with IMA Life. In the first phase, a feasibility study was carried out on the new Tile-X production line, comparing traditional computer vision techniques with deep learning approaches. The algorithms were developed using Halcon libraries within a dedicated environment for pharmaceutical applications. The results demonstrated that deep learning methods achieved higher accuracy and robustness while simplifying image preprocessing and system configuration. In the second phase, the selected algorithms were integrated into a deployable environment capable of real-time communication with the PLC and HMI, enabling fully automated inspection and counting. The resulting solution proved scalable, pharmaceutical-compliant, and effective in improving both the reliability and throughput of the packaging process. The developed programs successfully passed feasibility test and are scheduled for deployment on the production machine as part of the next development phase. Further work will focus on refining system performance under extended operating conditions and supporting future enhancements of the Tile-X line.