Methods, systems, and media for image searching
Methods, systems, and media for image searching are described. Images comprising one query image and a plurality of candidate images are received. For each candidate image, a first model similarity measure from an output of a first model configured for scene classification to perceive scenes in the images is determined. Further, for each candidate image of the plurality of candidate images, a second model similarity measure from the output of a second model configured for attribute classification to perceive attributes in the images is determined. For each candidate image of the plurality of candidate images, a similarity agglomerate index of a weighted aggregate of the first model similarity measure and the second model similarity measure is computed. The plurality of candidate images based on the respective similarity agglomerate index of each candidate image are ranked and a first ranked candidate images corresponding to the searched images are generated.
1. A method for image searching for images comprising at least one query image and a plurality of candidate images by ranking the plurality of candidate images based on similarity to the at least one query image, the method comprising:
determining, for each candidate image of the plurality of candidate images, a first model similarity measure from an output of a first model configured for scene classification to perceive scenes in the images, wherein the first model similarity measure is measured between each candidate image and the at least one query image;
determining, for each candidate image of the plurality of candidate images, a second model similarity measure from the output of a second model configured for attribute classification to perceive attributes in the images, wherein the second model similarity measure is measured between each candidate image and the at least one query image;
obtaining, for each candidate image of the plurality of candidate images, a similarity agglomerate index of a weighted aggregate of the first model similarity measure and the second model similarity measure;
ranking the plurality of candidate images based on the respective similarity agglomerate index of each candidate image and generating a first ranked candidate images corresponding to the searched images;
determining, for each candidate image of a subset of the first ranked candidate images, a third model similarity measure from the output of a third model configured for object detection, wherein the third model similarity measure is measured between each candidate image of the subset of the first ranked candidate images and the at least one query image;
ranking the subset of the first ranked candidate images based on the respective third model similarity measure of each candidate image of the subset of the first ranked candidate images and generating a second ranked candidate images corresponding to the searched images;
determining, for each candidate image of a subset of the second ranked candidate images, a fourth model similarity measure from the output of a fourth model configured for computing image statistics, wherein the fourth model similarity measure is measured between each candidate image of the subset of the second ranked candidate images and the at least one query image; and
ranking the subset of the second ranked candidate images based on the respective fourth model similarity measure and generating a final ranked candidate images corresponding to the searched images.
2. The method of claim 1 , wherein the fourth model is configured for computing the image statistics between the at least one query image and each candidate image of the subset of second ranked candidate images by:
transforming the at least one query image and each candidate image of the subset of the second ranked candidate images into at least one of a fixed length vector, histogram, and pixel values to generate at least one transformed query image and transformed candidate images; and
obtaining at least one of cosine similarity, hamming distance, L 0 distance, L 1 distance, and L 2 distance between the at least one transformed query image and each transformed candidate images.
3. The method of claim 1 , wherein the first model or the second model generates a classification confidence value for each class in the images, the first model similarity measure or the second model similarity measure being determined by the following equation
Similarity
measure
=
∑
class
∈
⋂
(
A
,
B
)
C
argument
(
⋂
(
A
,
B
)
)
∑
class
∈
⋃
(
A
,
B
)
C
max
(
⋃
(
A
,
B
)
)
where A and B are lists of the classes with respective classification confidence values in the at least one query image and each candidate image, respectively, ∩(A, B) is a class intersection list between A and B, ∩(A, B) comprising common classes between A and B, ∪(A, B) is a class union list between A and B, ∪(A, B) comprising unique classes of A and B, and C argument is the confidence value of one of the classes in the intersection list or the union list, wherein argument is a criterion of a minimum (min), a maximum (max), or an average (ave) confidence value of the one of the classes.
4. The method of claim 1 , wherein the first model or the second model generates a list of classes in the images, the first model similarity measure or the second model similarity measure being determined as in the following equation
Similarity
measure
=
#
⋂
(
A
,
B
)
#
⋃
(
A
,
B
)
where A and B are lists of classes in the at least one query image and each candidate image, respectively, #∩(A, B) is a number of classes in a class intersection list between A and B, #∩(A, B) comprising common classes between A and B, and #∪(A, B) is a number of classes in a class union list between A and 8 , #∪(A, B) comprising unique classes in A and B.
5. The method of claim 1 , wherein the third model configured for object detection comprises:
detecting objects in the image;
generating, for each object, an object label and a location;
obtaining, for each object, a relative area measurement as in the following equation
Relative
Area
object
=
Area
of
object
Area
of
image
where Area of object is a number of pixels represented by each object, and Area of image is a number of pixels of the image the object is detected in; and
determining the third model similarity measure as in the following equation
Third
model
similarity
measure
=
∑
object
label
∈
⋂
(
A
,
B
)
R
argument
(
⋂
(
A
,
B
)
)
∑
object
label
∈
⋃
(
A
,
B
)
R
max
(
⋃
(
A
,
B
)
)
where A and B are lists of object labels with respective relative area values in the at least one query image and each candidate image of the subset of the first ranked candidate images, respectively, ∩(A, B) is an object label intersection list between A and B, ∩(A, B) comprising common object labels between A and B, ∪(A, B) is an object label union list between A and B, ∪(A, B) comprising unique object labels in A and B, R argument is the relative area values of one of object labels in the object label intersection list or the object label union list, wherein argument is a criterion of a minimum (min), a maximum (max), and an average (ave) relative area value of the one of object labels.
6. The method of claim 5 , when A or B has more than one object of the one of object labels in the object label intersection list or object label union list, R argument is computed by:
determining R argument for every pair of the one of object labels in A and B, the pair being the one of object label in A and the one of object label in B;
including the relative area values of the one object label in A or B with no respective pair if R argument is computed for ∪(A, B); and
excluding the relative area values of the one object label A or B with no respective pair if R argument is computed for ∩(A, B).
7. The method of claim 6 , wherein computing R argument further comprises:
sorting the one of the object labels in A and B based on relative area values; and
ranking the one of the object labels in the sorted A and B.
8. The method of claim 5 , wherein the location of each object is represented by a mask or a bounding box.
9. A system for image searching for images comprising at least one query image and a plurality of candidate images by ranking the plurality of candidate images based on similarity to the at least one query image, the system comprising:
a non-transitory memory storing instructions;
a processor coupled to the memory and configured to execute the instructions to:
determine, for each candidate image of the plurality of candidate images, a first model similarity measure from an output of a first model configured for scene classification to perceive scenes in the images, wherein the first model similarity measure is measured between each candidate image and the at least one query image;
determine, for each candidate image of the plurality of candidate images, a second model similarity measure from the output of a second model configured for attribute classification to perceive attributes in the images, wherein the second model similarity measure is measured between each candidate image and the at least one query image;
obtain, for each candidate image of the plurality of candidate images, a similarity agglomerate index of a weighted aggregate of the first model similarity measure and the second model similarity measure;
rank the plurality of candidate images based on the respective similarity agglomerate index of each candidate image and generating a first ranked candidate images corresponding to the searched images;
determine, for each candidate image of a subset of the first ranked candidate images, a third model similarity measure from the output of a third model configured for object detection, wherein the third model similarity measure is measured between each candidate image of the subset of the first ranked candidate images and the at least one query image;
rank the subset of the first ranked candidate images based on the respective third model similarity measure of each candidate image of the subset of the first ranked candidate images and generating a second ranked candidate images corresponding to the searched images;
determine, for each candidate image of a subset of the second ranked candidate images, a fourth model similarity measure from the output of a fourth model configured for computing image statistics, wherein the fourth model similarity measure is measured between each candidate image of the subset of the second ranked candidate images and the at least one query image; and
rank the subset of the second ranked candidate images based on the respective fourth model similarity measure and generating a final ranked candidate images corresponding to the searched images.
10. The system of claim 9 , wherein the fourth model is configured for computing the image statistics between the at least one query image and each candidate image of the subset of second ranked candidate images by:
transforming the at least one query image and each candidate image of the subset of the second ranked candidate images into at least one of a fixed length vector, histogram, and pixel values to generate at least one transformed query image and transformed candidate images; and
obtaining at least one of cosine similarity, hamming distance, L 0 distance, L 1 distance, and L 2 distance between the at least one transformed query image and each transformed candidate images.
11. The system of claim 9 , wherein the first model or the second model generates a classification confidence value for each class in the images, the first model similarity measure or the second model similarity measure being determined by the following equation
Similarity
measure
=
∑
class
∈
⋂
(
A
,
B
)
C
argument
(
⋂
(
A
,
B
)
)
∑
class
∈
⋃
(
A
,
B
)
C
max
(
⋃
(
A
,
B
)
)
where A and B are lists of the classes with respective classification confidence values in the at least one query image and each candidate image, respectively, ∩(A, B) is a class intersection list between A and B, ∩(A, B) comprising common classes between A and B, ∪(A, B) is a class union list between A and B, ∪(A, B) comprising unique classes of A and B, and C argument is the confidence value of one of the classes in the intersection list or the union list, wherein argument is a criterion of a minimum (min), a maximum (max), or an average (ave) confidence value of the one of the classes.
12. The system of claim 9 , wherein the first model or the second model generates a list of classes in the images, the first model similarity measure or the second model similarity measure being determined as in the following equation
Similarity
measure
=
#
⋂
(
A
,
B
)
#
⋃
(
A
,
B
)
where A and B are lists of classes in the at least one query image and each candidate image, respectively, #∩(A, B) is a number of classes in a class intersection list between A and B, #∩(A, B) comprising common classes between A and B, and #(A, B) is a number of classes in a class union list between A and B, #∪(A, B) comprising unique classes in A and B.
13. The system of claim 9 , wherein the third model configured for object detection comprises:
detecting objects in the image;
generating, for each object, an object label and a location;
obtaining, for each object, a relative area measurement as in the following equation
Relative
Area
object
=
Area
of
object
Area
of
image
where Area of object is a number of pixels represented by each object, and Area of image is a number of pixels of the image the object is detected in; and
determining the third model similarity measure as in the following equation
Third
model
similarity
measure
=
∑
object
label
∈
⋂
(
A
,
B
)
R
argument
(
⋂
(
A
,
B
)
)
∑
object
label
∈
⋃
(
A
,
B
)
R
max
(
⋃
(
A
,
B
)
)
where A and B are lists of object labels with respective relative area values in the at least one query image and each candidate image of the subset of the first ranked candidate images, respectively, ∩(A, B) is an object label intersection list between A and B, ∩(A, B) comprising common object labels between A and B, ∪(A, B) is an object label union list between A and B, ∪(A, B) comprising unique object labels in A and B, R argument is the relative area values of one of object labels in the object label intersection list or the object label union list, wherein argument is a criterion of a minimum (min), a maximum (max), and an average (ave) relative area value of the one of object labels.
14. The system of claim 13 , when A or B has more than one object of the one of object labels in the object label intersection list or object label union list, R argument is computed by:
determining R argument for every pair of the one of object labels in A and B, the pair being the one of object label in A and the one of object label in B;
including the relative area values of the one object label in A or B with no respective pair if R argument is computed for ∪(A, B); and
excluding the relative area values of the one object label A or B with no respective pair if R argument is computed for ∩(A, B).
15. The system of claim 14 , wherein computing R argument further comprises:
sorting the one of the object labels in A and B based on relative area values; and
ranking the one of the object labels in the sorted A and B.