# US 10,504,209 B2 · Image dehazing method

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 10,504,209 B2
- Document ID: US10504209B2
- Status: Granted
- Kind: B2
- Application number: 15/448,546
- Filed: 2017-03-02
- Published: 2018-01-18
- Granted: 2019-12-10
- Earliest claimed: 2016-07-12
- Assignees: Beihang University
- Applicants: Beihang University
- Inventors: Xiaowu Chen; Yafei Song; Jia Li; Qinping Zhao; Xiaogang Wang
- Web page: https://patentgrep.com/patent/US10504209B2

## CPC

- G06F 18/2113: by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
- G06N 3/045: Combinations of networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06T 1/20: Processor architectures; Processor configuration, e.g. pipelining
- G06T 5/60: using machine learning, e.g. neural networks
- G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06V 10/764: using classification, e.g. of video objects
- G06V 10/771: Feature selection, e.g. selecting representative features from a multi-dimensional feature space
- G06V 10/82: using neural networks
- G06V 10/94: Hardware or software architectures specially adapted for image or video understanding
- G06F 18/2413: based on distances to training or reference patterns
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 20/20: Ensemble learning
- G06T 2200/28: involving image processing hardware
- G06T 2207/10024: Color image
- G06T 2207/20081: Training; Learning
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T 2207/30192: Weather; Meteorology
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

Present invention provides a ranking convolutional neural network constructing method and an image processing method and apparatus thereof. The ranking convolutional neural network includes a ranking layer that is configured to sort an output of a previous layer of the ranking layer, generate an output of the ranking layer according to the sorted output, and output the output of the ranking layer to a next layer of the ranking layer. Using the ranking convolutional neural network enables obtaining an output feature corresponding to the input feature image through automatic learning. Compared with prior art methods that obtain features through manual calculation, the method of the present invention is superior in terms of reflecting the objective laws contained by the patterns of the actual scene. When applied to the field of image processing, the method can significantly improve the effect of image processing.

## Claims

### Claim 1

An image dehazing method, comprising:
receiving an input hazy image;
determining, by using a ranking convolutional neural network, a feature vector for a current pixel of the input hazy image, wherein the ranking convolutional neural network comprises a ranking layer, a convolution layer, a downsampling layer and a fully-connecting layer, the ranking layer is configured to sort an output of a previous layer of the ranking layer, generate an output of the ranking layer according to the sorted output, and output the output of the ranking layer to a next layer of the ranking layer;
inputting the feature vector for the current pixel into a random forest regression model to obtain a haziness for the current pixel; and
determining a clear image corresponding to the input hazy image according to the haziness and ambient light information.

### Claim 2 (depends on 1)

The method according to claim 1, before the determining, by using a ranking convolutional neural network, a feature vector for a current pixel of the input hazy image, further comprising:
establishing correspondence between a feature vector of an output layer and the haziness;
wherein the correspondence between the feature vector of the output layer and the haziness is: for the feature vector of the output layer NB=(nB1, nB2, . . . , nBm), if tB∈(i/10−0.1,i/10], then nBi=1, otherwise nBi=0, wherein tB is the haziness.

### Claim 3 (depends on 1)

The method according to claim 1, wherein the ranking convolutional neural network comprises 10 layers that are, in order: an input layer, a first convolution layer, a first downsampling layer, the ranking layer, a second convolution layer, a third convolution layer, a second downsampling layer, a first fully-connecting layer, a second fully-connecting layer and an output layer, wherein a convolution kernel of the second convolution layer is smaller than a convolution kernel of the first convolution layer, and a convolution kernel of the third convolution layer is smaller than the convolution kernel of the first convolution layer;
the determining, by using a ranking convolutional neural network, a feature vector for a current pixel of the input hazy image comprises:
determining an input feature image for the current pixel through the input layer, wherein the input feature image is represented by a two dimensional matrix;
performing ranking, convoluting, downsampling and fully-connecting for the input feature image; and
taking a feature vector output by the second fully-connecting layer as the feature vector for the current pixel.

### Claim 4 (depends on 1)

The method according claim 1, wherein the inputting the feature vector for the current pixel into a random forest regression model to obtain a haziness for the current pixel comprises:
training a regression tree in the random forest regression model using a preset number of elements selected from the feature vector for the current pixel; obtaining an output value of the random forest regression model; and taking the output value of the random forest regression model as the haziness for the current pixel.

### Claim 5 (depends on 1)

The method according claim 1, after the determining a clear image corresponding to the input hazy image according to the haziness and ambient light information, further comprising:
adjusting brightness of the clear image using formula J*λJ, wherein
λ
=
log
(
∑
x
⁢
⁢
I
l
⁡
(
x
)
/
∑
x
⁢
⁢
J
l
⁡
(
x
)
)
+
1
,
I
l
⁡
(
x
)
is brightness of the input hazy image, and Jl(x) is the brightness of the clear image.