Deep computational holography
Techniques related to generating holographic images are discussed. Such techniques include application of a hybrid system including a pre-trained deep neural network and a subsequent iterative process using a sui table propagation model to generate diffraction pattern image data for a target holographic image such that the diffraction pattern image data is to generate a holographic image when implemented via a holographic display.
1. A system for generating holographic images comprising:
a memory to store target holographic image data; and
a processor coupled to the memory, the processor to:
convert the target holographic image data to multi-channel image data having an amplitude component channel and a phase component channel;
apply a pre-trained deep neural network to the multi-channel image data to generate intermediate diffraction pattern image data corresponding to the target holographic image data; and
apply an iterative process using a propagation model to the intermediate diffraction pattern image data to generate final diffraction pattern image data, wherein the propagation model corresponds to a holographic imaging arrangement to generate a holographic image using diffraction pattern image data, and wherein final modeled holographic image data generated by the iterative process using the propagation model and the final diffraction pattern image data meets an error threshold with respect to the target holographic image data and wherein intermediate modeled holographic image data generated based on the propagation model and the intermediate diffraction pattern image data does not meet the error threshold with respect to the target holographic image.
2. The system of claim 1 , wherein each pixel of the intermediate and final diffraction pattern image data comprises a normalized amplitude and wherein the holographic imaging arrangement comprises a spatial light phase modulator capable of modulating phase and incapable of modulating amplitude.
3. The system of claim 1 , wherein the pre-trained deep neural network is trained using a loss function comprising the propagation model and an error function corresponding to the error threshold and without use of ground truth diffraction pattern image data.
4. The system of claim 3 , wherein the error function comprises one of a peak signal to noise ratio error function, a minimum absolute error function, or a root mean square error function.
5. The system of claim 3 , wherein the pre-trained deep neural network is further trained using specialized training holographic image data corresponding to a predetermined application of the holographic imaging arrangement.
6. The system of claim 1 , wherein the multi-channel image data comprise a complex number representation having a real value representative of the amplitude component channel and an imaginary value representative of the phase component channel.
7. The system of claim 1 , wherein the processor to convert the target holographic image data comprises the processor to set each amplitude value of the amplitude component channel to a corresponding value of the target holographic image data and apply a phase pattern to generate phase values of the phase component channel.
8. The system of claim 7 , wherein the processor to apply the phase pattern comprises one of the processor to randomly generate the phase values or the processor to generate phase values to provide one or more regions of uniform amplitude based on the propagation model.
9. The system of claim 1 , wherein the pre-trained deep neural network comprises an encoder-decoder convolutional neural network having a plurality of convolutional layers each configured to apply one or more convolutional kernels.
10. The system of claim 1 , further comprising:
a holographic imaging device to present a final diffraction pattern image corresponding to the final diffraction pattern image data; and
a light source to project spatially coherent light onto the holographic imaging device to generate a final holographic image for a user.
11. The system of claim 1 , wherein the processor to apply the propagation model comprises at least one of the processor to apply a fast Fourier transform that models the holographic imaging arrangement or the processor to apply a quadratic phase profile and a fast Fourier transform that model the holographic imaging arrangement.
12. A method for generating holographic images comprising:
converting target holographic image data to multi-channel image data having an amplitude component channel and a phase component channel;
applying a pre-trained deep neural network to the multi-channel image data to generate intermediate diffraction pattern image data corresponding to the target holographic image data; and
applying an iterative process using a propagation model to the intermediate diffraction pattern image data to generate final diffraction pattern image data, wherein the propagation model corresponds to a holographic imaging arrangement to generate a holographic image using diffraction pattern image data, and wherein final modeled holographic image data generated by the iterative process using the propagation model and the final diffraction pattern image data meets an error threshold with respect to the target holographic image data and wherein intermediate modeled holographic image data generated based on the propagation model and the intermediate diffraction pattern image data does not meet the error threshold with respect to the target holographic image.
13. The method of claim 12 , wherein each pixel of the intermediate and final diffraction pattern image data comprises a normalized amplitude and wherein the holographic imaging arrangement comprises a spatial light phase modulator capable of modulating phase and incapable of modulating amplitude.
14. The method of claim 12 , wherein the pre-trained deep neural network is trained using a loss function comprising the propagation model and an error function corresponding to the error threshold and without use of ground truth diffraction pattern image data.
15. The method of claim 12 , wherein converting the target holographic image data comprises setting each amplitude value of the amplitude component channel to a corresponding value of the target holographic image data and applying a phase pattern to generate phase values of the phase component channel.
16. The method of claim 15 , wherein applying the phase pattern comprises one of randomly generating the phase values or generating phase values to provide one or more regions of uniform amplitude based on the propagation model.
17. The method of claim 12 , further comprising:
presenting a final diffraction pattern image corresponding to the final diffraction pattern image data via a holographic imaging device; and
projecting spatially coherent light onto the holographic imaging device to generate a final holographic image for a user.
18. At least one non-transitory machine readable medium comprising a plurality of instructions that, in response to being executed on a computing device, cause the computing device to generate holographic images by:
converting target holographic image data to multi-channel image data having an amplitude component channel and a phase component channel;
applying a pre-trained deep neural network to the multi-channel image data to generate intermediate diffraction pattern image data corresponding to the target holographic image data; and
applying an iterative process using a propagation model to the intermediate diffraction pattern image data to generate final diffraction pattern image data, wherein the propagation model corresponds to a holographic imaging arrangement to generate a holographic image using diffraction pattern image data, and wherein final modeled holographic image data generated by the iterative process using the propagation model and the final diffraction pattern image data meets an error threshold with respect to the target holographic image data and wherein intermediate modeled holographic image data generated based on the propagation model and the intermediate diffraction pattern image data does not meet the error threshold with respect to the target holographic image.
19. The non-transitory machine readable medium of claim 18 , wherein each pixel of the intermediate and final diffraction pattern image data comprises a normalized amplitude and wherein the holographic imaging arrangement comprises a spatial light phase modulator capable of modulating phase and incapable of modulating amplitude.
20. The non-transitory machine readable medium of claim 18 , wherein the pre-trained deep neural network is trained using a loss function comprising the propagation model and an error function corresponding to the error threshold and without use of ground truth diffraction pattern image data.
21. The non-transitory machine readable medium of claim 18 , wherein converting the target holographic image data comprises setting each amplitude value of the amplitude component channel to a corresponding value of the target holographic image data and applying a phase pattern to generate phase values of the phase component channel.
22. The non-transitory machine readable medium of claim 21 , wherein applying the phase pattern comprises one of randomly generating the phase values or generating phase values to provide one or more regions of uniform amplitude based on the propagation model.