Defects and components recognition in printed circuit boards using convolutional neural network

Abstract

This paper introduces an automated components recognition system for printed circuit boards using Convolutional Neural Network (CNN). In addition to that, localization on the defects of the PCB components is also presented. In the first stage, a simple convolutional neural network-based component recognition classifier is developed. Since training a convolutional neural network from scratch is expensive, transfer learning with pre-trained models is performed instead. Pre-trained models such as VGG-16, DenseNet169 and Inception V3 are used to investigate which model suits the best for components recognition. Using transfer learning with VGG-16, the best result achieved is 99% accuracy with the capability of recognizing up to 25 different components. Following that, object localization is performed using faster region-based convolutional neural network (R-CNN). The best mean average precision (mAP) achieved for the defects localization system is 96.54%.