Image quality assessment

1. A system comprising:

a first image sensor configured to capture a first image;

a second image sensor configured to capture a second image; and

a processing apparatus that is configured to:

receive the first image from the first image sensor;

receive the second image from the second image sensor;

stitch the first image and the second image to obtain a stitched image;

identify an image portion of the stitched image that is positioned on a stitching boundary of the stitched image;

input the image portion to a machine learning module to obtain a score, wherein the machine learning module has been trained using training data that included image portions labeled to reflect an absence of stitching and image portions labeled to reflect a presence of stitching, wherein the image portions labeled to reflect a presence of stitching included stitching boundaries of stitched images;

select a parameter of a stitching algorithm based at least in part on the score;

determining whether to re-stitch based on the score and a threshold;

responsive to determination, re-stitching, using the parameter, the first image and the second image to obtain a composite image; and

store, display, or transmit an output image based on the composite image.

2. The system of claim 1 , in which the processing apparatus is configured to:

identify one or more additional image portions within the stitched image that occur along the stitching boundary of the stitched image;

input the one or more additional image portions to the machine learning module to obtain one or more additional scores; and

select the parameter based on the one or more additional scores.

3. The system of claim 1 , in which the machine learning module comprises a feature extraction submodule that is configured to determine features based on the image portion.

4. The system of claim 1 , in which stitching to obtain the stitched image is performed such that individual pixels of the stitched image are respectively based on either the first image or the second image, but not both; and

in which stitching to obtain the composite image is performed such that at least one pixel of the composite image is based on both the first image and the second image.

5. The system of claim 1 , in which stitching to obtain the stitched image is performed such that individual pixels of the stitched image are respectively based on either the first image or the second image, but not both; and

in which the image portions of the training data labeled to reflect a presence of stitching included stitching boundaries of stitched images that were stitched without blending.

6. The system of claim 1 , in which the parameter specifies whether one dimensional parallax correction or two dimensional parallax correction will be applied to stich the first image and the second image.

7. The system of claim 1 , in which the image portion is a block of pixels from the stitched image that includes pixels on both sides of the stitching boundary, where the block of pixels has a resolution less than the resolution of the first image.

8. The system of claim 1 , in which the machine learning module includes a convolutional neural network.

9. The system of claim 1 , in which the machine learning module includes a neural network and the image portion is a block of pixels from the stitched image that includes pixels on both sides of the stitching boundary and all pixel values from the block of pixels are input to a first layer of the neural network.

10. A method comprising:

receiving a first image from a first image sensor;

receiving a second image from a second image sensor;

stitching the first image and the second image to obtain a stitched image;

identifying an image portion of the stitched image that is positioned on a stitching boundary of the stitched image; and

inputting the image portion to a machine learning module to obtain a score, wherein the machine learning module has been trained using training data that included image portions labeled to reflect an absence of stitching and image portions labeled to reflect a presence of stitching, wherein the image portions labeled to reflect a presence of stitching included stitching boundaries of stitched images;

select a parameter of a stitching algorithm based at least in part of the score;

determining whether to re-stitch based on the score and a threshold;

responsive to determination, re-stitching, using the parameter, the first image and the second image to obtain a composite image;

and store, display, or transmit an output image based on the composite image.

11. The method of claim 10 , comprising:

storing, displaying, or transmitting the score or a composite score based in part on the score.

12. The method of claim 10 , comprising:

identifying one or more additional image portions within the stitched image that occur along the stitching boundary of the stitched image;

inputting the one or more additional image portions to the machine learning module to obtain one or more additional scores; and

generating a histogram of the score and the one or more additional scores.

13. The method of claim 10 , comprising:

training the machine learning module, wherein the training data includes image portions detected with a single image sensor that are labeled to reflect an absence of stitching.

14. The method of claim 10 , comprising:

training the machine learning module, wherein the training data includes image portions labeled with subjective scores provided by humans for images from which the image portions are taken.

15. The method of claim 10 , comprising:

selecting a parameter of a stitching algorithm based on the score.

16. The method of claim 10 , in which the image portion is a block of pixels from the stitched image that includes pixels on both sides of the stitching boundary, where the block of pixels has a resolution less than the resolution of the first image.

17. The method of claim 10 , in which the machine learning module includes a neural network that receives pixel values from pixels in the image portion and outputs the score.

18. The method of claim 10 , comprising:

obtaining a plurality of scores from the machine learning module for a plurality of image portions from along the stitching boundary of the stitched image; and

determining a composite score for the stitched image based on the plurality of scores.