From Schematics to Netlists - Electrical Circuit Analysis Using Deep-Learning Methods

Within the project progressivKI, research is carried out to improve the analysis of schematics that depict an electrical circuit. Lots of manual efforts are necessary to validate a design, as schematics are handed in as image data. They neither follow a standard nor contain any meta information that can be obtained to automatically check certain conditions. Furthermore, even the visual representation of components like diodes, capacitors or resistors can differ depending on the design tool used.In this paper, we present an approach to decompose the problem into three different parts and describe their current status: (i) detection of the components like resistors, capacitors, or diodes (ii) detection of lines and their junctions (iii) detection of textual data placed next to components (like voltage or resistance). For each of the given areas we employ deep-learning methods as a basis. The training data is provided by Microchip in the form of link-annotated PDFs. In a preprocessing phase, the data is programmatically scanned for useful information like component names and bounding boxes to pre-annotate them before human correction. The final step is to fuse all information from (i)-(iii) to obtain a netlist that can be automatically validated with given rules.While most work has been carried out in (i) and (ii), a more general workflow including supportive tools has been established to extend our approach to PDFs from other design tools. The results show that recent deep-learning methods are capable of detecting components with a high accuracy given training data of good quality (no false labels).