Institutional Repository of Vilnius University
Hybrid system optimization for industrial deployment under limited data conditions /
Abstract
dc:descriptionThis bachelor thesis examines an industrial completeness inspection problem in which an Advent calendar moving on a conveyor belt must be checked to determine whether all compartments contain candy or chocolate. The work develops a hybrid computer vision system that combines OpenCV tray detection, tray normalization, compartment extraction via predefined zones, and a convolutional neural network that labels each compartment as filled or empty. The study is performed under limited data conditions. The original data comes from two factory videos that contain moving Advent calendars, empty frames, partially visible calendars, and frames that are unusable for reliable compartment extraction. Initial testing revealed very strong ResNet18 performance on curated compartment crops, but when it was used in a live demo, the same results could not be reproduced. Debugging showed that the main cause was not a weak classifier, but rather several preprocessing errors in the hybrid pipeline. Trays were not cropped correctly, an obsolete 180degree rotation remained in the normalization function, camera undistortion relied on incorrect frame resolution assumptions. Aſter these issues were resolved, the original ResNet18 system behaved correctly. The main conclusion is that reliability does not depend only on classifier accuracy, but it also depends on strict consistency between data preparation, zone alignment and the actual data passed to the CNN during inference.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Račkauskas, Olen,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308338563&prefLang=en_US