Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Assignees
Member
NoblisNoblis is a nonprofit research and technical organization supporting federal missions in defense, health, environment, and security. Emphasizing applied sciences, engineering, digital transformation, artificial intelligence, cloud, and cybersecurity, Noblis provides objective solutions for government agencies confronting complex operational and scientific challenges.
Noblis is a nonprofit research and technical organization supporting federal missions in defense, health, environment, and security. Emphasizing applied sciences, engineering, digital transformation, artificial intelligence, cloud, and cybersecurity, Noblis provides objective solutions for government agencies confronting complex operational and scientific challenges.
Publication Number
US-11275973-B2
Publication Date
2022-03-15
Expiration Date
Abstract
The present disclosure is directed to data classification. An exemplary computer-enabled method for classifying image data comprises: receiving an input image patch, wherein the input image patch is of a second data domain; providing the input image patch to a preprocessing algorithm to obtain a preprocessed image patch, wherein the preprocessing algorithm is trained to transform data of the second data domain to data of a first data domain; providing the preprocessed image patch to a pre-trained classification algorithm, wherein the pre-trained classification algorithm is trained based on training data of the first data domain; obtaining one or more classification outputs from the pre-trained classification algorithm based on the preprocessed image patch.
Core Innovation
The invention provides a computer-enabled method for classifying image data in which an input image patch is of a second data domain. The input image patch is provided to a preprocessing algorithm to obtain a preprocessed image patch, and the preprocessing algorithm is trained to transform data of the second data domain to data of a first data domain.
The preprocessed image patch is then provided to a pre-trained classification algorithm trained based on training data of the first data domain. The method obtains one or more classification outputs from the pre-trained classification algorithm based on the preprocessed image patch.
The disclosed embodiments further adapt the preprocessing algorithm to domain differences such as image style and image resolution range using a style/content preservation approach and resolution-range mapping. The preprocessing algorithm is trained using first and second image generators implemented in a modified CycleGAN framework with adversarial discriminator losses, cycle consistency loss, and identity mapping loss.
The disclosed training and generator embodiments include image generator architecture adjustments for resolution change and normalization choices that support resolution upscaling and tiling/large frames, including "down channel" normalization. The document also describes capturing data from the second data domain using one or more unmanned aircraft vehicles, and optionally associating the first data domain with ImageNet.
Claims Coverage
The document includes four independent claim sets: clm-00001 (method for classification), clm-00019 (electronic device executing classification pipeline), clm-00020 (non-transitory computer-readable storage medium implementing the pipeline), and clm-00021 (method for pre-processing to improve classification performance). The inventive features center on domain transformation preprocessing that maps a second data domain input to a first data domain representation before applying a pre-trained classifier trained on the first data domain.
Domain-transform preprocessing for second-to-first data domain classification
A computer-enabled method for classifying image data that receives an input image patch of a second data domain, provides it to a preprocessing algorithm to obtain a preprocessed image patch where the preprocessing algorithm is trained to transform data of the second data domain to data of a first data domain, provides the preprocessed image patch to a pre-trained classification algorithm trained based on training data of the first data domain, and obtains one or more classification outputs from the pre-trained classification algorithm based on the preprocessed image patch.
Executing the domain-transform preprocessing and classification pipeline on an electronic device
An electronic device comprising one or more processors, a memory, and one or more programs stored in the memory and configured to be executed, wherein the programs instruct the device to receive an input image patch of a second data domain, provide it to a preprocessing algorithm to obtain a preprocessed image patch where the preprocessing algorithm is trained to transform data of the second data domain to data of a first data domain, provide the preprocessed image patch to a pre-trained classification algorithm trained based on training data of the first data domain, and obtain one or more classification outputs based on the preprocessed image patch.
Non-transitory storage medium for domain-transform preprocessing and classification
A non-transitory computer-readable storage medium storing one or more programs comprising instructions which, when executed by one or more processors of an electronic device having a display, cause the electronic device to receive an input image patch of a second data domain, provide it to a preprocessing algorithm to obtain a preprocessed image patch where the preprocessing algorithm is trained to transform data of the second data domain to data of a first data domain, provide the preprocessed image patch to a pre-trained classification algorithm trained based on training data of the first data domain, and obtain one or more classification outputs from the pre-trained classification algorithm based on the preprocessed image patch.
Generating preprocessed image data by transforming second-domain inputs to first-domain for classification performance
A computer-enabled method for pre-processing image data to improve performance of classification algorithms comprising receiving data representing an input image of a second data domain, receiving data representing a pre-trained image classification algorithm trained based on a first data domain, and generating preprocessed image data based on the data representing the input image and the preprocessing algorithm, wherein the preprocessing algorithm is configured to transform data of the second data domain to data of a first data domain.
Across the independent claims, the core coverage is the domain-mapping preprocessing step that transforms second data domain image input into first data domain data before using a pre-trained image classification algorithm trained on first data domain training data, with corresponding implementations on an electronic device and as a non-transitory computer-readable storage medium, plus a separate independent preprocessing-focused claim that generates preprocessed image data to improve classification performance.
Stated Advantages
Documented Applications
No documented applications found
Interested in licensing this patent?