AI Molecular Structure Recognition With Node-Guided Bond Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional line connection-type object prediction devices face limitations in reducing errors and improving analysis accuracy and computation speed when predicting the state of objects connected by lines, particularly in structural formulas.
Innovation Solution
A line connection-type object prediction device using artificial intelligence that includes a processor to detect nodes and edges in molecular structural formulas through a backbone network, region proposal network, ROI module, and LOI module, with training based on node and edge labels, to output accurate prediction results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional line connection-type object prediction devices are used to predict the state of objects connected by lines in structural formulas, then the basic prediction function is provided, but prediction errors increase and analysis accuracy decreases
Solution Approach 1:
The patent segments the object prediction task into two distinct stages: node detection (identifying atomic positions) and edge detection (identifying bonds). This segmentation allows each stage to be optimized independently, with node detection using ROI modules and edge detection using LOI modules, thereby improving overall prediction accuracy while reducing errors in structural formula analysis
Solution Approach 2:
The patent introduces node detection information as an intermediary between the input image and edge detection. The detected node positions serve as mediators that guide the edge detection process, enabling more accurate bond identification by providing reference points for connecting atoms, thus improving prediction accuracy and reducing errors
2Productivity
If conventional prediction devices process structural formulas, then basic analysis is performed, but computation speed is slow
Solution Approach 1:
The patent performs node detection as a preliminary action before edge detection. By first identifying all node positions and using this information to guide subsequent edge detection, the system avoids redundant computations and focuses resources on identifying bonds between already-located atoms, thereby improving computation speed and reducing analysis time
Solution Approach 2:
The patent segments the computation into distinct node detection and edge detection phases, allowing each to be optimized independently. The node detection phase uses ROI pooling to efficiently process candidate regions, while the edge detection phase uses LOI modules to efficiently identify bonds based on node positions, improving overall computational efficiency
Data Source
AI summary
A line connection-type object prediction device and method using artificial intelligence in which objects are detected using atoms and bonds, which make up a molecular structural formula, as nodes and edges, respectively, when recognizing a molecular structural formula image representing the molecular structure of a compound, and the detection information about nodes is used when detecting edges for bonds.