Determination of population density using convoluted neural networks

Patent metadata

Field	Value
Publication number	US-10504007-B2
Application number	US-201715795741-A
Country	US
Kind code	B2
Filing date	Oct 27, 2017
Priority date	Oct 27, 2017
Publication date	Dec 10, 2019
Grant date	Dec 10, 2019

Abstract

In one embodiment, a method includes receiving an image on a computing device. The computing device may further execute a weakly-supervised classification algorithm to determine whether a target feature is present in the received image. As an example, the weakly-supervised classification algorithm may determine whether a building is depicted in the received image. In response to determining that a target feature is present, the method further includes using a weakly-supervised segmentation algorithm of the convoluted neural network to segment the received image for the target feature. Based on a determined footprint size of the target feature, a distribution of statistical information over the target feature in the image can be calculated.

First claim

What is claimed is:

1. A method comprising:
receiving an image;
executing a weakly-supervised classification algorithm to determine whether a target feature is present in the received image;
in response to determining that a target feature is present, using a weakly-supervised segmentation algorithm of the convoluted neural network to segment the received image for the target feature and determine a footprint size of the target feature; and
calculating a distribution of statistical information over the target feature based on the determined footprint size of the target feature.

2. The method of claim 1 , wherein calculating the distribution of statistical information is based at least in part on a property that scales with the size of the target feature within an image.

3. The method of claim 1 , wherein using the convoluted neural network to remove noise or haze comprises using an adaptive learnable transformer, wherein the adaptive learnable transformer is trained to remove noise or haze from pixels corresponding to the target feature.

4. The method of claim 3 , wherein training the adaptive learnable transformer is based on image-level labeled data and the weakly-supervised classification algorithm.

5. The method of claim 1 , wherein the weakly-supervised classification algorithm is trained using image-level labeled data, without pixel-level labeled data.

6. The method of claim 1 , wherein determining whether a target feature is present in the received image comprises:
for each pixel in the received image, determining, using the weakly-supervised classification algorithm, a per-pixel probability that the pixel corresponds to the target feature; and
determining an average of the per-pixel probabilities for the pixels in the received image.

7. The method of claim 1 , wherein the weakly-supervised classification algorithm further comprises a feedback loop to suppress irrelevant neuron activations of the convoluted neural network.

8. The method of claim 1 , wherein the weakly-supervised segmentation algorithm is trained using image-level labeled data, without pixel-level labeled data.

9. The method of claim 1 , wherein the weakly-supervised segmentation algorithm is trained to minimize a loss function

f
w

=

arg
⁢

⁢
min
⁢

∑
i

⁢

1
2

⁢



y
i

-

f
w

⁡

(

x
i

)



2

+

λ
⁢


w


2

wherein f w is the transformation from input x to output ŷ parametered with w.

10. The method of claim 1 , wherein the convoluted neural network comprises a plurality of layers.

11. The method of claim 10 , wherein each layer comprises a plurality of neurons.

12. The method of claim 11 , wherein for a particular layer 1 with input x 1 and target output y 1 , the convoluted neural network optimizes a target function min ½∥y l −f w (x l )∥ 2 +γ∥x l ∥1, wherein:
f w is the transformation from input x to output ŷ parametered with w; and
for a particular neuron n i l :

if
⁢

⁢

∂



y
l

-

f
w

⁡

(

x
l

)



2

∂

x
i
l

>

γ
⁢

⁢
then
⁢

⁢

x
i
l

⁢

⁢
may
⁢

⁢
be
⁢

⁢
positively
⁢

⁢
activated

;

if
⁢

⁢

∂



y

Assignees

• Facebook Inc

Inventors

• Tiecke Tobias Gerard
• Zhang Amy
• Liu Xianming
• Gros Andreas

Classifications

• G06V10/7753

  Incorporation of unlabelled data, e.g. multiple instance learning [MIL] · CPC title

• G06V10/82 Primary

  using neural networks · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/241

  relating to the classification model, e.g. parametric or non-parametric approaches · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title