Recognition of Electronic Component Orientations from Hand-Drawn Circuit Schematics through a Two Stage Machine Learning System

Large strides in Artificial Intelligence (AI) and Machine Learning (ML) have lead to the automation of countless manual and time intensive tasks in circuit analysis and simulation. As hand-drawing circuits is often a critical time-consuming step, automating the netlisting and schematic-generation will greatly speedup this manual process. With emerging traction towards AI, research is increasingly being conducted in this field to develop component detection and classifier models. In this paper, a novel two stage detection system with a focus on component accuracy and orientations is developed. Using a hand-drawn circuit dataset of 2304 images by various authors with different drawing styles, an object-detection model was developed using Faster-RCNN, and YOLOv5 to recognize circuit components. Through a custom created dataset of 84 orientation classes for 15 types of electronic components, a secondary classification model was developed using ResNet-50 for the recognition of orientation information with an accuracy of 99.97%.