Back to results
US 12,361,099 B2Granted
Method and apparatus for creating high-fidelity, synthetic data for artificial intelligence model training and inference
AssigneeCignal LLC
InventorEric M. Fiterman
Life of this patent
Estimating the term…
Abstract
New software systems and capabilities that facilitate the rapid development, evaluation, and deployment of advanced inspection and detection systems using high-fidelity synthetic imagery. The present invention generates high-fidelity synthetic imagery for detection systems that analyze data across the electromagnetic spectrum in an automated, random, directed, or semi-directed manner.
Claims15
An apparatus for generating synthetic data for artificial intelligence models, comprising:
- at least one processor;
- at least one non-transient computer memory;
- an object library stored on said at least one non-transient computer memory including a catalog of data representing a first set of shapes, objects, and/or geometries;
- a data management module stored on said at least one non-transient computer memory and containing computer readable instructions executable by said at least one processor for loading and saving data files representing a second set of shapes, objects, and geometries, and converting said data files into optimized data structures;
- a materials library stored on said at least one non-transient computer memory including a catalog of data representing a plurality of materials and associated data representing electromagnetic wave behavior of said plurality of materials, for one or more simulated energies, wherein said plurality of materials comprises liquids, gases and solids;
- a scene management module for 1) selection of a first object from said catalog of data representing shapes, objects and/or geometries, 2) virtually placing into an interior volume of said first object one or more second objects from said catalog of data representing said second set of shapes, objects, and/or geometries and 3) selecting a material from said plurality of materials and associating said selected material with at least said second object, and 4) generating a plurality of three-dimensional scenes comprising said selected first objects with said selected interior second objects, and said selected material and associated electromagnetic wave behavior data, including object position, internal composition and motion of said first and second objects, and electromagnetic wave behavior for said first and second objects and said selected materials in the scene, the module stored on said at least one non-transient computer memory and containing computer instructions executable by said at least one processor for storing compressed representations of said first and second objects, said selected materials and said electromagnetic wave behavior into scene data configured to be uncompressed and loaded at run-time during simulation;
- an energy simulator stored on said at least one non-transient computer memory and containing computer readable instructions executable by said at least one processor for generating synthetic data for a particular scene using said associated electromagnetic wave behavior properties, said synthetic data for said particular scene reflecting electromagnetic wave behavior properties for said second objects and said selected materials in said interior volumes of said first objects;
- a data generation module stored on said at least one non-transient computer memory and containing computer readable instructions executable by said at least one processor for converting said synthetic data for said particular scene into a data output format, and
- a display interface module stored on said at least one non-transient computer memory and containing computer readable instructions executable by said at least one processor for creating, displaying, and interacting with said data output format based on user entries.
6 more elements
A method for generating synthetic imagery for training artificial intelligence models for security and screening applications, comprising:
- building/assembling/storing a volumetric object library comprising firearms, explosives, incendiaries, narcotics, controlled substances, enclosures, vehicles, structures, animals, and human figures using a data management module that loads object data from a variety of 2D and 3D data as slices of real or virtual objects, surfaces, meshes, voxels, points, paths, lines, curves, 3D models, Computer-aided Design (CAD) files, or specified procedural geometries;
- storing a material library to comprising a plurality of materials that may be virtually placed in interior volumes of items in said volumetric object library;
- selecting one or more volumetric objects from the object library and selecting one or more materials from the material library, virtually placing one or more said materials in said interior volumes of said volumetric objects, and generating a variety of scenes, wherein a scene management module packs or places objects into one or more enclosures randomly or using computerized instructions as part of a packing algorithm defined in a scene generation specification and wherein the scene management module is configured to use object layer data to alter material properties randomly or using specific algorithms;
- loading and populating a scene from the volumetric objects in a scene file definition into an energy simulation engine configured to handle the placement, orientation, sizing, motion, materials, and deformations from the scene file definition;
- performing simulation of energy waves by the energy simulation engine according to a supplied job configuration that details, wherein users may specify parameters regarding radiation energy, energy binning, frequency, amplitude, wavelength, propagation, absorption, reflection, and refraction of electromagnetic waves and particles with various bodies in a scene.
2 more elements
Dependencies are inferred from the claim text.