Label-free non-reference image quality assessment via deep neural network

Patent metadata

Field	Value
Publication number	US-2016379352-A1
Application number	US-201514931843-A
Country	US
Kind code	A1
Filing date	Nov 3, 2015
Priority date	Jun 24, 2015
Publication date	Dec 29, 2016
Grant date	—

Abstract

A method for training a neural network to perform assessments of image quality is provided. The method includes: inputting into the neural network at least one set of images, each set including an image and at least one degraded version of the image; performing comparative ranking of each image in the at least one set of images; and training the neural network with the ranking information. A neural network and image signal processing tuning system are disclosed.

First claim

What is claimed is:

1 . A method for training a neural network to perform assessments of image quality, the method comprising:
inputting into the neural network at least one set of images, each set comprising an image and at least one degraded version of the image; performing comparative ranking of each image in the at least one set of images; and training the neural network with the ranking information.

2 . The method as in claim 1 , wherein the comparative ranking is performed in a comparative layer.

3 . The method as in claim 2 , wherein the comparative layer implements a sigmoid function to provide pairwise ranking of the images within each set of images.

4 . The method as in claim 3 , wherein the sigmoid function comprises:

h


(

y
i

,

y
j

)

=

1

1
+



l

i
·
j



(

y
i

-

y
j

)

wherein, y i and y j represent output quality scores associated with input images, x i and x j , respectively; and
and l i,j represents prior information for pairwise ranking of y i and y j output by the comparative layer.

5 . The method as in claim 4 , wherein learning rules for the comparative layer comprise:

∂

∂

y
i

=

-

l

i
·
j



1

1
+



l

i
·
j



(

y
i

-

y
j

)



(

1
-

1

1
+



l

i
·
j



(

y
i

-

y
j

)

)

;

∂

∂

y
j

=

l

i
·
j



1

1
+



l

i
·
j



(

y
i

-

y
j

)



(

1
-

1

1
+



l

i
·
j

Assignees

• Samsung Electronics Co Ltd

Inventors

• Zhang Qiang
• Ji Zhengping
• Shi Lilong
• Ovsiannikov Ilia

Classifications

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06N3/084 Primary

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

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06T2207/20008

  Globally adaptive · CPC title