ElectroNet: An Enhanced Model for Small-Scale Object Detection in Electrical Schematic Diagrams

Abstract

This study presents the design and development of a highly accurate and efficient approach for interpreting both hand-drawn and printed electric-circuit schematics. It addresses critical challenges in the field such as the lack of annotated data forobject-detection in schematics. To mitigate this, we have created an extensive dataset, comprising of 23 classes of electric-circuit elements, textual elements with units, prefixes, decimal points, and an array of various handwriting and printing styles with a total of 198495 annotations in the dataset. This rich and diverse dataset ensures the robustness of our system against different styles of circuit drawings. We combined our dataset with the MNIST dataset to enhance the model’s resilience to diverse writing styles. Another obstacle faced in this domain is the accurate recognition and assignment of textual labels in schematics. To overcome this, we employed object-detection methodologies, leveraging use of hybrid model. It has proven effective in analyzing line diagrams and incorporating various language OCRs. Despite initial success with minor parameter adjustments, a shortfall in the accuracy of object-detection was identified when significant size variations occurred between objects. To address this, we trained a new model, leading to significant improvements in the recognition of circuit elements and textual symbols across all sizes, with high precision. This enhanced system showed an overall correct simulation rate of 92%, and an overall correct recognition rate of 94% for textual symbols. This work establishes the efficacy of employing object-detection techniques for interpreting electric schematics while also underlining the challenges and potential avenues for future research in this domain. Our approach has shown notable improvements in recognizing electric-circuit drawings, particularly for smaller components, and opens avenues for further advancements in the field of circuit recognition and analysis.

Parsing Netlists of Integrated Circuits from Images via Graph Attention Network

“…Subsequently, with the development of deep learning, there have been studies integrating it into the recognition of hand-drawn electrical schematics [ 4 , 5 , 6 ]. The most recent papers can be traced back to 2015, when De et al [ 7 ] proposed a circuit image segmentation algorithm.…”