End-to-end Digitization of Handwritten Circuit Diagrams

Abstract

Handwritten engineering diagrams are still commonly used in many industries and academic settings, but the lack of digitization limits their utility in modern workflows. While significant effort has been made to digitize handwritten content in other engineering domains, the digitization of handwritten circuit diagrams remains underexplored. Despite the growing demand for automated digitization systems in electronics and electrical engineering, there remains a lack of comprehensive datasets and research efforts targeting this specific domain. This thesis addresses these challenges through contributions to both dataset preparation and model development for an end-to-end digitization system. A dataset titled Digitize-HCD was developed, comprising 1,277 handwritten circuit diagrams from over 150 volunteers. The dataset includes detailed annotations for component symbol recognition, text label recognition, and component port localization, featuring 18,602 component symbol annotations across 17 classes and 11,936 text labels comprising 44,443 characters. Additionally, a component port localization dataset was created using Gaussian heatmaps to represent port locations for each component type. Four baseline object detection models—RTMDet, YOLOv8, Faster R-CNN, and EfficientDet—were tested for component symbol detection, with YOLOv8 (CSPDarknet P5 backbone) achieving the highest overall mAP of 79.9%, and an mAP50 of 98.9%. YOLOv8 was selected as the final model due to its balance of high accuracy, low computational cost (FLOPs of 129G), and low latency (54.8 ms). Text detection was performed using the Differentiable Binarization Network (DBNet), while recognition employed the Show, Attend and Read (SAR) framework. A U-Netbased framework using Gaussian heatmap predictions was applied for component port localization. With this framework, simpler components like Resistors (MSE = 1.697, SDR = 98.89%) and Inductors (MSE = 1.586, SDR = 99.12%) showed high accuracy, while more complex components such as MOSFET (N-Channel) (MSE = 3.776, SDR = 89.19%) demonstrated higher prediction errors. The results from the component symbol detection, text detection and recognition, and port localization modules were integrated to reconstruct the circuit topology and generate a SPICE-compatible netlist, enabling seamless simulation of digitized handwritten circuits. This research contributes a publicly accessible dataset and a comprehensive end-to-end system for the end-to-end digitization of handwritten circuit diagrams, capable of transforming handwritten circuit diagrams into machine readable formats.