63 International Conference on Advanced Technologies (ICAT’18)
E-ISBN: 978-605-68537-0-8 Antalya, Turkey, 28 April -May 1, 2018 DEEP LEARNING BASED APPROACHES FOR MACHINE VISION
INSPECTION APPLICATIONS
MEHMET BAYGIN1, MEHMET KARAKOSE2
1 Ardahan University, Turkey; 2 Firat University, Turkey
ABSTRACT
Today, machine vision applications are frequently used in many different areas such as automotive, food, textile, and electronics. Especially with these systems that allow contactless control, many different industrial fields can produce products with no fault or near-fault accuracy. In contactless measurement systems, products that pass over a conveyor are usually controlled via one or more cameras. At this point, the defective products are separated on the conveyor and the products are delivered without error to the end user. In this paper, a new machine vision application has been developed which can perform non-contact measurement. The control and separation of labels produced in an industrial production line has been performed with the proposed approach in the study. For this purpose, the posts of the packaging labels on the wrapped position on a rotating cylinder is checked and the system can distinguish it if any mistakes are faulted. The installed system uses one camera and can take about 50 fps. Deep learning method is used as the basis for this approach proposed in the study. In the system, images taken primarily from the camera are subjected to various image preprocessing steps. The clean images obtained after this process are stored for use in the training phase of the deep learning. After enough images are obtained for the training phase, these images are sent to the training module of the deep learning system. After completion of the training phase, test images are taken from the system and compared with the training model. As a result of the comparison process, the system divides the image frames taken for the test into two classes, correct and faulty. When the results obtained from the experimental studies are examined, it has been observed that the deep learning-based machine vision approach gives very accurate and effective results.
