TLDR: The paper introduces Image2Net, a hybrid framework that converts image-based analog circuit diagrams into text-based netlists, which are crucial for training Large Language Models (LLMs) in circuit design. It also provides new diverse datasets and a novel evaluation metric called Netlist Edit Distance (NED), achieving significantly higher accuracy and lower error rates compared to previous methods.
Designing analog integrated circuits (ICs) has traditionally been a complex task, heavily relying on the experience and manual effort of engineers. While automation could significantly speed up this process, traditional tools often fall short due to the intricate nature of analog circuits. Large Language Models (LLMs) have shown immense promise in various fields, including digital circuit design, thanks to their ability to understand and generalize knowledge. However, when it comes to analog ICs, LLMs face a major hurdle: a severe lack of suitable data.
The core issue is that most existing analog circuits are represented as image-based circuit diagrams, not the text-based netlists that LLMs need for training. This data gap prevents LLMs from effectively learning and generating analog circuit designs. Previous attempts to convert these diagrams into netlists have been limited, struggling with the wide variety of image styles and circuit elements found in real-world designs. Furthermore, existing evaluation methods for these conversions often lack the precision needed to truly assess accuracy, sometimes misclassifying incorrect results as correct.
To address these critical challenges, a new research paper introduces a comprehensive solution called Image2Net. This innovative work not only proposes a hybrid framework for converting circuit diagrams to netlists but also provides new, diverse datasets and a more accurate evaluation metric.
New Datasets for Better Training
A significant contribution of this research is the release of new open-source datasets. These datasets are designed to overcome the limitations of previous ones, which often featured simplified drawing styles or incomplete circuit diagrams. The Image2Net datasets include:
-
Device Identification Dataset: Comprising 2914 complete circuit images with diverse styles and complexities, featuring over 48,000 annotated devices.
-
Crossing Identification Dataset: Annotates over 28,000 crossings in nearly 2000 images, classifying them into three types (Bridge, Flat, Dot) to understand how wires connect or cross without connecting.
-
Device Orientation Classification Dataset: Provides data for determining the orientation (up, right, down, left) and mirror information of various devices like MOS, BJT, Diode, and AMP devices, crucial for accurate netlist generation.
-
Netlist Evaluation Dataset: A validation set of 104 manually verified circuit diagrams and their corresponding netlists, ranging from simple to complex, provided in JSON format.
These datasets are richer in style and complexity, derived from textbooks, academic papers, and the internet, ensuring a more robust training environment for conversion algorithms.
Introducing Netlist Edit Distance (NED) for Precise Evaluation
To accurately measure the performance of circuit diagram-to-netlist conversion, the researchers introduced a new metric called Netlist Edit Distance (NED). Unlike traditional accuracy or Graph Edit Distance (GED) which can be insufficient, NED precisely assesses the correctness of connections while ignoring irrelevant factors like device or network names. It calculates the minimum number of operations (insertions, deletions, substitutions of nodes or edges) required to transform a converted netlist’s graph into the ground truth graph, then normalizes this value. A lower NED indicates a more accurate conversion, providing a refined indicator of recognition effect.
The Image2Net Framework: A Hybrid Approach
The Image2Net framework itself employs a hybrid methodology, combining neural networks with a series of computer vision (CV) algorithms. The process involves several key steps:
-
Object Detection: Uses YOLO-V8 to identify the positions and types of all components and crossings in the circuit diagram.
-
Wire Identification: Employs binarization and skeletonization algorithms to detect wires, removing distracting elements like text.
-
Jumper Identification: Determines whether a crossing represents a jumper (wires crossing without connecting) or a true connection, treating jumpers as virtual devices.
-
Port Positioning: Identifies the intersection points between component bounding boxes and wires, and uses MobileNetV2 to classify device orientation and mirror information, establishing the correct port order.
-
Line Tracking and Post-Processing: Connects ports that intersect the same wire domain and performs post-processing to merge networks and eliminate virtual devices, ultimately generating the final netlist.
This innovative approach allows Image2Net to handle multi-style and complex circuit recognition effectively.
Also Read:
- Bridging AI Reasoning and Engineering Principles for Automated Analog Circuit Design
- AIGer: A New Deep Learning Model for Enhanced Logic Circuit Design
Impressive Results
The Image2Net framework demonstrates significant improvements over previous works. Based on their benchmark, Image2Net achieved an 80.77% successful rate in converting circuit diagrams to netlists, which is 34.62% to 45.19% higher than previous methods. More specifically, it showed an averaged NED of 0.116, which is 62.1% to 69.6% lower than state-of-the-art solutions. This indicates a much higher accuracy in capturing the correct circuit topology and connections.
This research marks a crucial step forward in automating analog IC design. By effectively converting complex, image-based circuit diagrams into machine-readable netlists, Image2Net enriches the knowledge available to Large Language Models, paving the way for LLMs to gain a deeper understanding of analog circuits and contribute more significantly to their design. For more details, you can read the full research paper here.
