Prototype Development of an Automated Vision-Based IV Bag Inspection System for Pharmaceutical Quality Control.

This thesis presents the development of a prototype vision-based system for the automated inspection of visible particulate contamination in flexible intravenous (IV) bags. The work addresses the problem as an integrated combination of optical design, image acquisition, computer vision, classification, and prototype development. Controlled particle samples with different optical properties were prepared and used to evaluate several illumination approaches. Brightfield and darkfield configurations were retained as complementary solutions, since different particle materials produced different contrast mechanisms. An experimental acquisition station was then developed to generate short image sequences after controlled product agitation. The computer-vision pipeline processes consecutive frames through image differencing, candidate extraction, temporal association, and track-grid generation. The resulting representations are classified into three classes: bubble, particle, and meniscus. Two ResNet-50-based classification workflows were evaluated. The HALCON implementation achieved 97.27% accuracy on the labeled test set, while the Python TensorFlow/Keras implementation achieved 92.42%. The results confirm the feasibility of the proposed optical and vision methodology while also showing that classifier performance alone does not represent complete inspection performance. The thesis concludes with the ongoing development of a mechanically reproducible station. The integrated prototype has been demonstrated experimentally, while final mechanical reproducibility and complete system-level validation remain under development.