Defect classification of electronic boards by deep learning

Within the scope of this thesis, a study was carried out to detect production defects in printed circuit boards using deep learning. Studies on this subject in the literature generally focus on defects due to missing components and soldering errors. In this thesis, a situation assessment was made based on the case of Vestel Beyaz Eşya Sanayi, and in addition to missing components and soldering errors, a study was also carried out to detect reverse polarity and misalignment defects, which are relatively less studied in the literature. In this context, a total of 783 images of defective and defect-free PCBs obtained from various production lines of Vestel. The number of these images was increased to 1407 using various data augmentation techniques. Afterwards, a YOLOv5 model was builded for object detection with deep learning. Model performance was evaluated through precision, sensitivity, mean average precision and intersection over union metrics and the following performances are achieved: precison [0.694-0.912], recall [0.857-0.877], mean average precision [0.865-0.931] and intersection over union [0.735-0.786].