Target detection

1. A target detection method, comprising:

classifying, using at least a first classification model, a plurality of image patches included in an input image, to obtain one or more candidate image patches, in the plurality of image patches, that are preliminarily classified as including a target; and

for each candidate image patch:

extracting a corresponding salience area;

constructing a corresponding target feature vector based, at least in part, on the corresponding salience area; and

classifying, using at least a second classification model, the corresponding target feature vector to determine whether the candidate image patch includes the target.

2. The method according to claim 1 ,

wherein the first classification model includes a deep neural network, and the deep neural network includes an input layer, at least one convolutional layer, and at least one fully connected layer, and

wherein extracting a corresponding salience area comprises:

generating a class activation map for the candidate image patch based, at least in part, on a feature map that is for the candidate image patch and that is output by the last convolutional layer in the at least one convolutional layer; and

obtaining, based, at least in part, on the class activation map for the candidate image patch, the corresponding salience area for the candidate image patch.

3. The method according to claim 2 ,

wherein obtaining, based, at least in part, on the class activation map for the candidate image patch, the corresponding salience area for the candidate image patch comprises:

binarizing the class activation map for the candidate image patch, to obtain a binarized class activation map for the candidate image patch, the corresponding salience area being a connected area in the binarized class activation map.

4. The method according to claim 3 ,

wherein obtaining, based, at least in part, on the class activation map for the candidate image patch, the corresponding salience area for the candidate image patch further comprises:

upsampling, before the binarizing is performed, the class activation map for the candidate image patch, so that an upsampled class activation map has the same size as the candidate image patch,

wherein the binarizing is performed on the upsampled class activation map.

5. The method according to claim 2 , wherein generating a class activation map for the candidate image patch based, at least in part, on a feature map that is for the candidate image patch and that is output by the last convolutional layer in the at least one convolutional layer comprises:

generating the class activation map for the candidate image patch by performing global average pooling on the feature map that is for the candidate image patch and that is output by the last convolutional layer in the at least one convolutional layer.

6. The method according to claim 3 ,

wherein at least some of the plurality of image patches overlap, and

wherein constructing a corresponding target feature vector comprises:

constructing the corresponding target feature vector to include at least one of:

a salience area proportion feature, representing a relative size of the connected area in the binarized class activation map for the candidate image patch;

a salience area number feature, representing the number of connected areas in the binarized class activation map for the candidate image patch;

a salience area hitting rate feature, representing the number of binarized class activation maps for the candidate image patch that include the connected area in the binarized class activation map for the candidate image patch at the same time; and

a first classification model output feature, representing a probability that the candidate image patch includes the target.

7. The method according to claim 6 , wherein for a j th candidate image patch, the salience area proportion feature is expressed as:

ar

j

=

1

-

salience

⁢

⁢

area

patch

⁢

⁢

area

,

wherein salience area=Σ k=1 h Σ l=1 h v kl , v kl is a pixel value of a pixel in a k th row and an l th column in a binarized class activation map for the candidate image patch, the candidate image patch is an h×h image patch, and patch area=h 2 is the area of the candidate image patch.

8. The method according to claim 6 , wherein for a j th candidate image patch, the salience area number feature is expressed as:

sn

j

=

1

-

1

salience

⁢

⁢

number

,

wherein salience number is the number of connected areas in a binarized class activation map for the candidate image patch.

9. The method according to claim 6 , wherein for a j th candidate image patch, the salience area hitting rate feature is expressed as:

ht

j

=

1

-

1

hitting

⁢

⁢

times

,

wherein hitting times is the total number of all binarized class activation maps for the candidate image patch that include connected areas in a binarized class activation map for the candidate image patch.

10. The method according to claim 6 , wherein for a j th candidate image patch, the first classification model output feature is expressed as p j , wherein p j represents a probability that the candidate image patch includes the target.

11. The method according to claim 1 , wherein classifying, using at least a first classification model, a plurality of image patches included in an input image, to obtain one or more candidate image patches, in the plurality of image patches, that are preliminarily classified as including a target comprises:

for each of the plurality of image patches, outputting, by the first classification model, a corresponding probability that the image patch includes the target; and

determining an image patch, in the plurality of image patches, with its corresponding probability greater than a threshold, as one of the one or more candidate image patches.

12. The method according to claim 1 , wherein the plurality of image patches are obtained from the input image by using a sliding window, and at least some of the plurality of image patches overlap.

13. The method according to claim 12 , wherein before the plurality of image patches are classified using at least the first classification model, all of the plurality of image patches are obtained by sliding the sliding window on the input image.

14. The method according to claim 1 , further comprising:

performing non-maximum suppression on a candidate image patch that is determined using at least the second classification model as including the target, to obtain a detection result of all targets included in the input image.

15. A computing device, comprising:

one or more processors; and

a memory having instructions stored thereon, wherein the instructions, when executed by the one or more processors, cause the one or more processors to perform actions comprising:

classifying, using at least a first classification model, a plurality of image patches included in an input image, to obtain one or more candidate image patches, in the plurality of image patches, that are preliminarily classified as including a target; and

for each candidate image patch:

extracting a corresponding salience area;

constructing a corresponding target feature vector based, at least in part, on the corresponding salience area; and

classifying, using at least a second classification model, the corresponding target feature vector to determine whether the candidate image patch includes the target.

16. A non-transitory computer-readable storage medium having instructions stored thereon, wherein the instructions, when executed by one or more processors, cause the one or more processors to perform a method comprising:

classifying, using at least a first classification model, a plurality of image patches included in an input image, to obtain one or more candidate image patches, in the plurality of image patches, that are preliminarily classified as including a target; and

for each candidate image patch:

extracting a corresponding salience area;

constructing a corresponding target feature vector based, at least in part, on the corresponding salience area; and

classifying, using at least a second classification model, the corresponding target feature vector to determine whether the candidate image patch includes the target.