1. A method (400) of correcting glare, the method (400) comprising:
receiving (402), by a processor (104), a plurality of image frames captured by an
imaging device; and
determining (404), by the processor (104), a glare-corrected image frame corresponding
to each of the plurality of image frames in real-time by using an unsupervised autoencoder
neural network model,
wherein the unsupervised autoencoder neural network model is trained based
on a ground truth image frame determined for each of the plurality of image frames, and
wherein the ground truth image frame for an image frame of the plurality of
image frames is determined by:
determining (408), by the processor (104), a grey-scale image frame of
a corresponding image frame from the plurality of image frames;
converting (410), by the processor (104), the grey-scale image frame to
a frequency domain image frame by performing a Fourier transformation;
determining (412), by the processor (104), a peak frequency in the
frequency domain image frame;
determining (414), by the processor (104), a filtered image frame by
attenuating the peak frequency in the frequency domain image frame using a notch filter;
determining (416), by the processor (104), a filtered spatial domain
image from the filtered image frame by performing an inverse Fourier transformation; and
determining (422), by the processor (104), the corresponding ground
truth image frame based on the filtered spatial domain image and the corresponding image
frame.
2. The method (400) as claimed in claim 1, wherein the determination of the ground truth image
frame for each of the plurality of image frames comprises:
determining (418), by the processor (104), a mask based on thresholding the filtered
spatial domain image in order to determine pixels corresponding to the peak frequency in the
corresponding image frame;
determining (420), by the processor (104), one or more masked pixels in the filtered
spatial domain image using the mask; and
determining (424), by the processor (104), the corresponding ground truth image frame
by replacing one or more pixels corresponding to the peak frequency in the corresponding
17
image frame with the corresponding one or more masked pixels in the filtered spatial domain
image.
3. The method (400) as claimed in claim 1, wherein the thresholding of the filtered spatial
domain image is performed based on a predefined threshold level.
4. The method (400) as claimed in claim 1, wherein the ground truth image frame for each of
the plurality of image frames is determined in real-time.
5. The method (400) as claimed in claim 4, wherein the unsupervised autoencoder neural
network model is trained in real-time based on the ground truth image frame determined for
each of the plurality of image frames.
6. A system (100) for correcting glare in real-time, comprising:
a processor (104); and
a memory (106) communicably coupled to the processor (104), wherein the memory
(106) stores processor-executable instructions, which, on execution, cause the
processor (104) to:
receive a plurality of image frames captured by an imaging device and;
determine a glare-corrected image frame corresponding to each of the plurality
of image frames in real-time by using an unsupervised autoencoder neural network
model,
wherein the unsupervised autoencoder neural network model is trained
based on a ground truth image frame determined for each of the plurality of image
frames, and
wherein to determine the ground truth image frame for each of the
plurality of image frames, the processor (104) is configurable to:
determine a grey-scale image frame of a corresponding image
frame from the plurality of image frames;
convert the grey-scale image frame to a frequency domain image
frame by performing a Fourier transformation;
determine a peak frequency in the frequency domain image
frame;
18
determine a filtered image frame by attenuating the peak
frequency in the frequency domain image frame using a notch filter;
determine a filtered spatial domain image from the filtered image
frame by performing an inverse Fourier transformation; and
determine the corresponding ground truth image frame based on
the filtered spatial domain image and the corresponding image frame.
7. The system (100) as claimed in claim 6, wherein to determine the ground truth image frame
for each of the plurality of image frames, the processor (104) is configurable to:
determine a mask based on thresholding the filtered spatial domain image in order to
determine pixels corresponding to the peak frequency in the corresponding image frame;
determine one or more masked pixels in the filtered spatial domain image using the
mask; and
determine the corresponding ground truth image frame by replacing one or more pixels
corresponding to the peak frequency in the corresponding image frame with the corresponding
one or more masked pixels in the filtered spatial domain image.
8. The system (100) as claimed in claim 6, wherein the thresholding of the filtered spatial
domain image is performed based on a predefined threshold level.
9. The system (100) as claimed in claim 6, wherein the ground truth image frame for each of
the plurality of image frames is determined in real-time.
10. The system (100) as claimed in claim 9, wherein the unsupervised autoencoder neural
network model is trained in real-time based on the ground truth image frame determined for
each of the plurality of image frames.