Defect Classification of Printed Circuit Boards based on Transfer Learning

In this paper, we propose a transfer learning based method to classify PCB images into two classes - i) True Defect and ii) Pseudo Defect A pre-trained Inception-V3 model is used for transfer learning and the mid-level representations of the PCB images are extracted from the output of an intermediate layer of the pre-trained model. These mid-level representations of the PCB images are used to train a small adaptation network, which makes the overall network suitable for classification of PCB images. In order to tackle overfitting, regularization strategies are implemented. The results of our experiment, conducted on real world PCB images, show a significant improvement in the classification accuracy as compared to previous work.