Fusion and super-resolution of monochrome and color images using deep neural networks
One embodiment of the present invention sets forth a technique for combining a color image and a monochrome image. The technique includes inputting a first set of pixel values from the color image and a second set of pixel values from the monochrome image into a neural network. The technique also includes applying, via a first set of neural network layers, one or more downsampling operations and one or more upsampling operations to feature maps associated with the inputted pixel values to generate a first output image that includes a first set of color values for a first set of pixel locations in the monochrome image. The technique further includes applying a second set of neural network layers to the first set of color values to generate a second output image that includes a second set of color values that is larger than the first set of color values.
1 . A computer-implemented method for combining a color image and a monochrome image, the method comprising:
inputting a first set of pixel values from the color image and a second set of pixel values from the monochrome image into a neural network;
applying, via a first set of neural network layers included in the neural network, one or more downsampling operations and one or more upsampling operations to feature maps associated with the first set of pixel values, the second set of pixel values, and a third set of pixel values to generate a first output image, wherein the first output image includes a first set of color values for a first set of pixel locations in the monochrome image, wherein the first set of pixel values, second set of pixel values, and third set of pixel values are captured by a respective first sensor, second sensor, and third sensor, wherein each set of pixel values are taken of a same scene and at a similar time;
and
applying a second set of neural network layers included in the neural network to the first set of color values to generate a second output image that includes a second set of color values, wherein the second set of color values is larger than the first set of color values.
2 . The computer-implemented method of claim 1 , wherein the third sensor is a red green blue (RGB) sensor.
3 . The computer-implemented method of claim 1 , further comprising determining at least one of the first set of pixel values or the second set of pixel values based on an alignment of the color image and the monochrome image.
4 . The computer-implemented method of claim 1 , further comprising:
synthesizing the color image and the monochrome image based on a first target image; and
training the neural network based on one or more losses between the second output image and the first target image.
5 . The computer-implemented method of claim 4 , further comprising:
applying the neural network to a fourth set of pixel values captured by a monochrome sensor and a fifth set of pixel values captured by an RGB sensor to generate a third output image; and
training the neural network based on one or more additional losses between the third output image and a second target image captured by an additional RGB sensor.
6 . The computer-implemented method of claim 4 , wherein the one or more losses comprise an L1 loss between the second set of color values and a third set of color values from the first target image.
7 . The computer-implemented method of claim 1 , wherein the second set of neural network layers comprises at least one of a convolutional layer or a pixel shuffle layer.
8 . The computer-implemented method of claim 1 , wherein the second output image is associated with a higher resolution than the first output image.
9 . The computer-implemented method of claim 1 , wherein the neural network includes a U-Net architecture.
10 . One or more non-transitory computer-readable media storing instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of:
inputting a first set of pixel values from a color image and a second set of pixel values from a monochrome image into a neural network;
applying, via a first set of neural network layers included in the neural network, one or more downsampling operations and one or more upsampling operations to feature maps associated with the first set of pixel values, the second set of pixel values, and a third set of pixel values, to generate a first output image, wherein the first output image includes a first set of color values for a first set of pixel locations in the monochrome image; and
applying a second set of neural network layers included in the neural network to the first set of color values to generate a second output image that includes a second set of color values, wherein the second set of color values is larger than the first set of color values, wherein the first set of pixel values, second set of pixel values, and third set of pixel values are captured by a respective first sensor, second sensor, and third sensor, wherein each set of pixel values are captured of a same scene and at a similar time.
11 . The one or more non-transitory computer-readable media of claim 10 , wherein the instructions further cause the one or more processors to perform the step of applying the neural network to additional feature maps associated with one or more additional images to produce the first output image.
12 . The one or more non-transitory computer-readable media of claim 11 , wherein the one or more additional images comprise an additional color image that temporally precedes the color image and an additional monochrome image that temporally precedes the monochrome image.
13 . The one or more non-transitory computer-readable media of claim 10 , wherein the instructions further cause the one or more processors to perform the steps of:
applying the neural network to a fourth set of pixel values captured by a monochrome sensor and a fifth set of pixel values captured by an RGB sensor to generate a third output image; and
training the neural network based on one or more losses between the third output image and a first target image captured by an additional RGB sensor.
14 . The one or more non-transitory computer-readable media of claim 13 , wherein the instructions further cause the one or more processors to perform the steps of:
synthesizing the color image and the monochrome image based on a second target image; and
training the neural network based on one or more additional losses between the second output image and the second target image.
15 . The one or more non-transitory computer-readable media of claim 13 , wherein the additional RGB sensor is associated with a higher resolution than the RGB sensor.
16 . The one or more non-transitory computer-readable media of claim 10 , wherein the second set of neural network layers comprises at least one of a convolutional layer and a pixel shuffle layer.
17 . The one or more non-transitory computer-readable media of claim 10 , wherein the third sensor is an RGB sensor.
18 . A wearable device, comprising:
a monochrome sensor;
an RGB sensor;
one or more memories that store instructions, and
one or more processors that are coupled to the one or more memories and, when executing the instructions, are configured to:
determine, based on an alignment of a color image captured by the RGB sensor and a monochrome image captured by the monochrome sensor, a first set of pixel values from the color image and a second set of pixel values from the monochrome image;
apply, via a first set of neural network layers, one or more downsampling operations and one or more upsampling operations to feature maps associated with the first set of pixel values, the second set of pixel values, and a third set of pixel values, to generate a first output image, wherein the first output image includes a first set of color values for a first set of pixel locations in the monochrome image; and
apply a second set of neural network layers to the first set of color values to generate a second output image that includes a second set of color values, wherein the second set of color values is larger than the first set of color values, wherein the third set of pixel values is captured by a second RGB sensor and wherein each set of pixel values are taken of a same scene and at a similar time.