Semantic refinement of image regions
Examples are described of segmenting an image into image regions based on depicted categories of objects, and for refining the image regions semantically. For example, a system can determine that a first image region in an image depicts a first category of object. The system can generate a color distance map of the first image region that maps color distance values to each pixel in the first image region. A color distance value quantifies a difference between a color value of a pixel in the first image region and a color value of a sample pixel in a second image region in the image. The system can process the image based on a refined variant of the first image region that is refined based on the color distance map, for instance by removing pixels from the first image region whose color distances fall below a color distance threshold.
1. An apparatus for image processing, the apparatus comprising:
at least one memory; and
at least one processor coupled to the at least one memory, the at least one processor configured to:
receive an image captured by an image sensor;
determine a first image region in the image that includes a depiction of a first object category of a plurality of object categories and a second image region in the image that includes a depiction of a second object category of the plurality of object categories;
combine information derived from each of a plurality of color distance maps to generate a decision map, wherein each of the plurality of color distance maps includes a plurality of color distance values indicative of respective differences in color between a sample pixel in the second image region and each of a plurality of pixels in the first image region, and wherein each of the plurality of color distance maps is based on a different sample pixel in the second image region; and
process the image based on a refined variant of the first image region that is refined based on the decision map.
2. The apparatus of claim 1 , the at least one processor configured to:
identify, for each one of the plurality of color distance maps, a subset of the plurality of color distance values that fall below a predetermined color distance threshold to obtain the information, wherein the decision map indicates a subset of pixels in the first image region that are to be removed from the first image region,
wherein refining the first image region includes removing the subset of the pixels in the first image region from the first image region as indicated by the decision map.
3. The apparatus of claim 2 , the at least one processor configured to:
add the subset of the pixels that are removed from the first image region to the second image region to generate a refined variant of the second image region; and
process the image also based on the refined variant of the second image region to process the image based on the refined variant of the first image region.
4. The apparatus of claim 2 , the at least one processor configured to:
identify sample pixels that are included in a particular image region, wherein the particular image region is one of either a third image region or a fourth image region, wherein the second image region includes the third image region and the fourth image region, wherein the third image region includes a depiction of the second object category of the plurality of object categories, wherein the fourth image region includes a depiction of a third object category of the plurality of object categories;
wherein the decision map is generated based on the subset of the plurality of color distance maps that correspond to the sample pixels included in the particular image region, and
process the image based also on a refined variant of the particular image region to process the image based on the refined variant of the first image region, wherein the subset of the pixels that are removed from the first image region are added to the particular image region to generate the refined variant of the particular image region.
5. The apparatus of claim 1 , the at least one processor further configured to:
identify one or more clusters of pixels within the second image region based on a clustering algorithm; and
wherein different sample pixels are selected from different clusters and the selection is based on a position of the sample pixel relative to at least one other pixel in the cluster from which the sample pixel is selected.
6. The apparatus of claim 1 , the at least one processor further configured to:
identify one or more clusters of pixels within the second image region based on a clustering algorithm; and
wherein different sample pixels are selected from different clusters and the selection is based on a proximity of the sample pixel to a boundary between the first image region and the second image region.
7. The apparatus of claim 1 , the at least one processor further configured to:
process the refined variant of the first image region of the received image using a first setting and processing at least a portion of the second image region of the received image using a second setting distinct from the first setting.
8. The apparatus of claim 7 , wherein both the first setting and the second setting are tuning settings of an image signal processor (ISP), and wherein processing the received image based on the refined variant of the first image region includes processing the received image using the ISP and using the tuning settings of the ISP.
9. The apparatus of claim 1 , wherein, to process the received image based on the refined variant of the first image region, the at least one processor is configured to apply post-processing to at least the refined variant of the first image region of the received image.
10. The apparatus of claim 1 , the at least one processor further configured to:
downscale the received image to generate a downscaled instance of the, wherein determining that the first image region includes the depiction of the first object category includes identifying the depiction of the first object category within the downscaled instance of the received image.
11. The apparatus of claim 1 , wherein determining that the first image region includes the depiction of the first object category is based on an output of a trained neural network, the trained neural network generating the output in response to receiving, as input, one of the received image and a downscaled copy of the received image.
12. The apparatus of claim 1 , the at least one processor further configured to:
generate a category map of the received image identifying a plurality of image regions that includes the first image region and the second image region, wherein each of the plurality of image regions corresponds to a different object category of the plurality of object categories, and wherein processing the received image is based on the category map.
13. The apparatus of claim 12 , the at least one processor is configured to:
generate a confidence map of the received image mapping each pixel of the received image to a confidence based on a confidence in determining that said each pixel maps to a particular object category of the plurality of object categories in the category map, wherein processing the received image is further based on the confidence map.
14. The apparatus of claim 13 , wherein, to process the received image based on the refined variant of the first image region, the at least one processor is further configured to apply a processing setting at a particular strength for each pixel of the received image, wherein the particular strength is based on the particular confidence level of said each pixel in the confidence map.
15. The apparatus of claim 1 , wherein the apparatus is one of a mobile device, a wireless communication device, a head-mounted display, and a camera.
16. The apparatus of claim 1 , further comprising the image sensor.
17. The apparatus of claim 1 , wherein the at least one processor includes at least one of an image signal processor (ISP) and a classification engine.
18. The apparatus of claim 1 , further comprising a display configured to display the received image after the at least one processor processes the received image.
19. A method of image processing, the method comprising:
receiving an image captured by an image sensor;
determining a first image region in the image that includes a depiction of a first object category of a plurality of object categories and a second image region in the image that includes a depiction of a second object category of the plurality of object categories;
combining information derived from each of a plurality of color distance maps to generate a decision map, wherein each of the plurality of color distance maps includes a plurality of color distance values indicative of respective differences in color between a sample pixel in the second image region and each of a plurality of pixels in the first image region, and wherein each of the plurality of color distance maps is based on a different sample pixel in the second image region; and
processing the image based on a refined variant of the first image region that is refined based on the decision map.
20. The method of claim 19 , further comprising:
identifying, for each one of the plurality of color distance maps, a subset of the plurality of color distance values that fall below a predetermined color distance threshold to obtain the information, wherein the decision map indicates a subset of pixels in the first image region that are to be removed from the first image region,
wherein refining the first image region includes removing the subset of the pixels in the first image region from the first image region as indicated by the decision map.
21. The method of claim 20 , further comprising:
adding the subset of the pixels that are removed from the first image region to the second image region to generate a refined variant of the second image region; and
processing the image also based on the refined variant of the second image region to process the image based on the refined variant of the first image region.
22. The method of claim 20 , further comprising:
identifying sample pixels that are included in a particular image region, wherein the particular image region is one of either a third image region or a fourth image region, wherein the second image region includes the third image region and the fourth image region, wherein the third image region includes a depiction of the second object category of the plurality of object categories, wherein the fourth image region includes a depiction of a third object category of the plurality of object categories;
wherein the decision map is generated based on the subset of the plurality of color distance maps that correspond to the sample pixels included in the particular image region, and
processing the image based also on a refined variant of the particular image region to process the image based on the refined variant of the first image region, wherein the subset of the pixels that are removed from the first image region are added to the particular image region to generate the refined variant of the particular image region.
23. The method of claim 19 , further comprising:
identifying one or more clusters of pixels within the second image region based on a clustering algorithm; and
wherein different sample pixels are selected from different clusters and the selection is based on a position of the sample pixel relative to at least one other pixel in the cluster from which the sample pixel is selected.
24. The method of claim 19 , wherein processing the image based on the refined variant of the first image region includes processing the refined variant of the first image region of the received image using a first setting and processing at least a portion of the second image region of the received image using a second setting distinct from the first setting.
25. The method of claim 24 , wherein both the first setting and the second setting are tuning settings of an image signal processor (ISP), and wherein processing the image based on the refined variant of the first image region includes processing the image using the ISP and using the tuning settings of the ISP.
26. The method of claim 19 , wherein processing the image based on the refined variant of the first image region includes applying post-processing to at least the refined variant of the first image region of the received image.
27. The method of claim 19 , further comprising:
downscaling the received image to generate a downscaled instance of the, wherein determining that the first image region includes the depiction of the first object category includes identifying the depiction of the first object category within the downscaled instance of the received image.
28. The method of claim 19 , wherein determining that the first image region includes the depiction of the first object category is based on an output of a trained neural network, the trained neural network generating the output in response to receiving, as input, one of the received image and a downscaled copy of the received image.
29. The method of claim 19 , further comprising:
generating a category map of the received image identifying a plurality of image regions that includes the first image region and the second image region, wherein each of the plurality of image regions corresponds to a different object category of the plurality of object categories, and wherein processing the received image is based on the category map.
30. The method of claim 29 , further comprising:
generating a confidence map of the received image mapping each pixel of the received image to a confidence based on a confidence in determining that said each pixel maps to a particular object category of the plurality of object categories in the category map, wherein processing the received image is further based on the confidence map.