Image Classification using Deep Learning and Machine Learning Approach on FICS-PCB image Dataset

Abstract

Printed Circuit Boards (PCBs) provide functional support to many electronic devices such as laptops, biomedical devices and other critical devices by connecting electrical compo nents, traces, and vias on the board. However, these connected circuits may contain defects trace shortages, malicious component replacements, etc., that can impact the system’s in tegrity. Specific examples include but are not limited to functional failure of the device like sensor failure in a self-driving car , user data leakage , or partial or full system control taken by adversaries. Therefore, it is critical to inspect PCBs before they are deployed. Existing PCB inspection methods fall into two categories i.e. electrical testing and auto mated visual inspection (PCB-AVI). Electrical testing methods cannot detect malicious im plants that alter PCB functions outside the tested locations and automated visual inspection are evaluated based on only private datasets which makes performance comparison between methods difficult. To address this, in this report we use FICS-PCB dataset which consists of PCB images featuring multiple types of components and various image conditions to fa cilitate performance evaluation in challenging scenarios that are likely to be encountered in practice. The electronic components are very small in size and the images in that dataset for individual components are taken after high zooming which can induce noise in the images. We evaluate various noise-filtering techniques and found that bilateral-filtering performs the best. Considering the limited number of images of particular components, we use data augmentation techniques such as rotation and flipping of images to increase the size of training dataset and we perform transfer learning. Finally, we evaluate machine and deep-learning models on the dataset and observe that for our work ResNet-34 provides the highest classification accuracy of 94.93% for a certain number of training epochs.