Optical imaging
AssigneeOxford University Innovation Limited
InventorsGiorgio Maltese, Anastasiia Pushkina, Josje da Costa Filho, Alexander Lvovsky
Life of this patent
Estimating the term…
Abstract
A method (100) of training an image processing unit (27) for use in optical imaging, the method (100) comprising: providing (102) a training object (37); computing (104) a plurality of theoretically expected signals generated by detecting the component of the electromagnetic field arriving from the object (37) in each of a plurality of different spatial modes of light; and generating (106) a reconstructed image (39) based on the theoretically expected signals, wherein the reconstructed image (39) is provided as a label for the training object (37) for use in training the image processing unit (27).
Claims20
A computer implemented method of generating training data for use in training a trainable image processing algorithm implemented on an image processor to be used in optical imaging, the method comprising:
- for each spatial mode of a plurality of different spatial modes of light, detecting a component of light reflected or emitted by a first training object in that mode;
- simulating an expected output of the image processor when imaging the first training object by:
- for each spatial mode of the plurality of different spatial modes of light, determining a theoretically expected signal that is measured by detecting the component of light reflected or emitted by the training object in that mode;
- and
- generating an image based on the determined theoretically expected signals in the plurality of different spatial modes of light;
- and
- associating the detected components of light with the generated image, the generated image being a label for the training object, wherein a variance between the label and an image constructed by using the trainable image processing algorithm to process the detected components of light suitable is used in training the trainable image processing algorithm.
A method of training an image processor for use in optical imaging, the image processor configured to use a trainable image processing algorithm to process light reflected or emitted by an object in each spatial mode of a plurality of different spatial modes of light to generate an image of the object, the method comprising:
- generating a set of training data by
- providing a plurality of training objects, and
- for each training object,
- detecting the component of light reflected or emitted by the training object in each spatial mode of the plurality of different spatial modes of light;
- for each spatial mode of the plurality of different spatial modes of light, determining a theoretically expected signal that is measured by detecting the component of light reflected or emitted by the training object;
- and
- generating an image based on the determined theoretically expected signals in the plurality of different spatial modes of light;
- and
- associating the detected components of light with the generated image, the generated image being a label for the training object;
- using the image processor to construct detected images of the plurality of training objects by using the trainable image processing algorithm to process the detected components of light;
- comparing the images constructed by the image processor to the associated label for each training object to determine a variance between the image constructed by the image processor and the associated label;
- and
- updating the trainable image processing algorithm based on the determined variance between the image constructed by the image processor and the associated label.
Dependencies are inferred from the claim text.