Image crop
A saliency map of an image is generated. Saliency regions of the saliency map are identified. The saliency regions are merged into a combined saliency region. Candidate image crops of the image are generated based on the combined saliency region. An image crop of the image is selected from the candidate image crops using a machine learning model.
1 . A non-transitory computer-readable data storage medium storing program code executable by a processor to perform processing comprising:
generating a saliency map of an image;
identifying a plurality of saliency regions of the saliency map;
merging the saliency regions into a combined saliency region;
generating a plurality of candidate image crops of the image based on the combined saliency region; and
selecting an image crop of the image from the candidate image crops using a machine learning model, wherein the machine learning model is a neural network trained as a twin neural network based on reference images and image crops of the reference images using a ranking loss objective in which the image crops are negative samples and the reference images are positive samples.
2 . The non-transitory computer-readable data storage medium of claim 1 , wherein generating the saliency map from the image comprises using a deep supervised machine learning model having a skip-layer structure.
3 . The non-transitory computer-readable data storage medium of claim 1 , wherein identifying the saliency regions of the saliency map comprises:
performing a connected-component analysis on the saliency map to segment the saliency map into the saliency regions.
4 . The non-transitory computer-readable data storage medium of claim 3 , wherein identifying the saliency regions of the saliency map further comprises:
discarding any saliency region having an area less than a threshold percentage of an area of the saliency map.
5 . The non-transitory computer-readable data storage medium of claim 1 , wherein merging the saliency regions into the combined saliency region comprises:
identifying a bounding box of the saliency map encompassing the saliency regions.
6 . The non-transitory computer-readable data storage medium of claim 1 , wherein generating the candidate image crops based on the combined saliency region comprises:
dividing the image into a grid of grid regions;
specifying a first sub-grid of the grid regions at a first corner of the image;
specifying a second sub-grid of the grid regions at a second corner of the image diagonally opposite the first corner; and
identifying each candidate image crop that has a first corner in any grid region of the first sub-grid and a second corner in any grid region of the second sub-grid, that has an area greater than a threshold percentage of an area of the image, and that has an aspect ratio within a specified range, and that fully covers the combined saliency region.
7 . The non-transitory computer-readable data storage medium of claim 1 , wherein generating the candidate image crops based on the combined saliency region comprises:
adjusting the combined saliency region to have a specified aspect ratio;
successively enlarging the combined saliency region horizontally and vertically while maintaining the specified aspect region until the combined saliency region horizontally and/or vertically exceeds the image, to specify a plurality of enlargements of the combined saliency region;
at each enlargement of the combined saliency region, cropping the image in correspondence with the combined saliency region to identify one of the candidate image crops; and
discarding any candidate image crop having an area less than a threshold percentage of an area of the image.
8 . The non-transitory computer-readable data storage medium of claim 7 , wherein generating the candidate image crops based on the combined saliency region further comprises:
repeating, for each of one or multiple additional specified aspect ratios, adjusting the combined saliency region, successively enlarging the combined saliency region until the combined saliency region horizontally and/or vertically exceeds the image, at each enlargement cropping the image, and discarding any candidate image having an area less than the threshold percentage of the area of the image.
9 . The non-transitory computer-readable data storage medium of claim 1 , wherein selecting the image crop of the image from the candidate image crops using the machine learning model comprises:
inputting each candidate image crop into the machine learning model and receiving as output from the machine learning model a corresponding crop score for each candidate image crop; and
selecting the image crop of the image as the candidate image crop having a highest corresponding crop score received as output from the machine learning model.
10 . A method comprising:
generating a plurality of image crops for each of a plurality of reference images;
training a neural network as a twin neural network based on the reference images and the image crops using a ranking loss objective in which the image crops are negative samples and the reference images are positive samples; and
selecting an image crop of an image from a plurality of candidate image crops of the image using the trained neural network.
11 . The method of claim 10 , wherein the ranking loss objective is based on a distance between each image crop and a corresponding reference image.
12 . The method of claim 11 , wherein the ranking loss objective is equal to a maximum of zero and a margin minus the distance.
13 . The method of claim 10 , further comprising:
generating the candidate image crops based on a saliency map of the image.
14 . A computing device comprising:
a processor; and
memory storing instructions executable by the processor to:
generate a saliency map of an image using a first machine learning model;
generate a plurality of candidate image crops based on the saliency map; and
select an image crop of the image from the candidate image crops using a second machine learning model, wherein the second machine learning model is a neural network trained as a twin neural network based on reference images and image crops of the reference images using a ranking loss objective in which the image crops are negative samples and the reference images are positive samples.
15 . The method of claim 13 , wherein the saliency map is generated using a deep supervised machine learning model having a skip-layer structure.
16 . The computing device of claim 14 , wherein the first machine learning model is a deep supervised machine learning model having a skip-layer structure.
17 . The computing device of claim 14 , wherein instructions are executable by the processor to generate the plurality of candidate image crops based on the saliency map by:
dividing the image into a grid of grid regions;
specifying a first sub-grid of the grid regions at a first corner of the image;
specifying a second sub-grid of the grid regions at a second corner of the image diagonally opposite the first corner; and
identifying each candidate image crop that has a first corner in any grid region of the first sub-grid and a second corner in any grid region of the second sub-grid, that has an area greater than a threshold percentage of an area of the image, and that has an aspect ratio within a specified range, and that fully covers the saliency map.
18 . The computing device of claim 14 , wherein instructions are executable by the processor to generate the plurality of candidate image crops based on the saliency map by:
adjusting the saliency map to have a specified aspect ratio;
successively enlarging the saliency map horizontally and vertically while maintaining the specified aspect region until the saliency map horizontally and/or vertically exceeds the image, to specify a plurality of enlargements of the saliency map;
at each enlargement of the saliency map, cropping the image in correspondence with the saliency map to identify one of the candidate image crops; and
discarding any candidate image crop having an area less than a threshold percentage of an area of the image.
19 . The computing device of claim 18 , wherein instructions are executable by the processor to generate the plurality of candidate image crops based on the saliency map by further:
repeating, for each of one or multiple additional specified aspect ratios, adjusting the saliency map, successively enlarging the saliency map until the saliency map horizontally and/or vertically exceeds the image, at each enlargement cropping the image, and discarding any candidate image having an area less than the threshold percentage of the area of the image.