Image authenticity detection method and apparatus

Patent metadata

Field	Value
Publication number	US-12488567-B2
Application number	US-202217979883-A
Country	US
Kind code	B2
Filing date	Nov 3, 2022
Priority date	May 11, 2021
Publication date	Dec 2, 2025
Grant date	Dec 2, 2025

Abstract

This disclosure is directed to an image authenticity detection method and apparatus. The method includes: obtaining an image; removing low-frequency information from the image to obtain first image information of the image; denoising the first image information to obtain second image information; determining, based on a difference between the first image information and the second image information, a fixed pattern noise feature map corresponding to the image; analyzing distribution of fixed pattern noise in the fixed pattern noise feature map, the fixed pattern noise being inherent noise from a camera sensor and not interfered by image content; and detecting, based on the distribution, authenticity of the image to obtain an authenticity detection result of the image.

First claim

What is claimed is:

1 . An image authenticity detection method, performed by a computer device, comprising:
obtaining an image; removing low-frequency information from the image to obtain first image information of the image; denoising the first image information to obtain second image information; determining, based on a difference between the first image information and the second image information, a fixed pattern noise feature map corresponding to the image; analyzing distribution of fixed pattern noise in the fixed pattern noise feature map, the fixed pattern noise being inherent noise from a camera sensor and not interfered by image content; and detecting, based on the distribution of fixed pattern noise in the fixed pattern noise feature map, authenticity of the image to obtain an authenticity detection result of the image, the authenticity detection result including that the image is an authentic image or the image is a forged image, wherein the detecting further comprises in response to the distribution matching fixed pattern noise distribution of a preset forgery type, determining the image to be a forged image of the forgery type based on the forgery type corresponding to the matched fixed pattern noise distribution, the forgery type comprises at least one of a type of tampering, a type of synthesis, or a type of artificial intelligence generation.

2 . The method according to claim 1 , wherein the first image information comprises wavelet coefficients of high-frequency components in a plurality of directions of the image in wavelet domain, and the removing the low-frequency information from the image to obtain the first image information comprises:
performing wavelet transform on the image in spatial domain to decompose the image into a low-frequency component and high-frequency components in a plurality of directions in the wavelet domain; and setting a wavelet coefficient of the low-frequency component to zero to obtain wavelet coefficients of the high-frequency components in the plurality of directions.

3 . The method according to claim 2 , wherein the second image information comprises denoised high-frequency wavelet coefficients corresponding to the high-frequency components in the plurality of directions of the image in the wavelet domain, and the denoising the first image information to obtain denoised second image information comprises:
denoising the wavelet coefficients of the high-frequency components in the plurality of directions to obtain the denoised high-frequency wavelet coefficients corresponding to the high-frequency components in the plurality of directions.

4 . The method according to claim 3 , wherein the obtaining the fixed pattern noise feature map corresponding to the image comprises:
obtaining high-frequency noise wavelet coefficients corresponding to the high-frequency components in the plurality of directions based on a difference between the wavelet coefficients of the high-frequency components and the denoised high-frequency wavelet coefficients of the high-frequency components that correspond to a same direction; and performing inverse wavelet transform based on the wavelet coefficient of the low-frequency component that is set to zero and the high-frequency noise wavelet coefficients corresponding to the high-frequency components in the plurality of directions, to obtain the fixed pattern noise feature map corresponding to the image.

5 . The method according to claim 4 , wherein the performing the wavelet transform on the image in spatial domain comprises:
performing multi-scale wavelet transform on the image in the spatial domain to decompose the image into a low-frequency component, and high-frequency components at a plurality of scales, high-frequency components at a same scale comprising high-frequency components in a plurality of directions.

6 . The method according to claim 4 , wherein the performing the inverse wavelet transform to obtain the fixed pattern noise feature map corresponding to the image comprises:
performing inverse wavelet transform based on the wavelet coefficient of the low-frequency component that is set to zero and the high-frequency noise wavelet coefficients corresponding to the high-frequency components in the plurality of directions at the plurality of scales, to obtain the fixed pattern noise feature map corresponding to the image.

7 . The method according to claim 3 , wherein the denoising the wavelet coefficients of the high-frequency components in the plurality of directions to obtain the denoised high-frequency wavelet coefficients corresponding to the high-frequency components in the plurality of directions comprises:
estimating local variances of non-noise information in the high-frequency components in the plurality of directions; and for each of the high-frequency components in each of the plurality of directions, performing noise filtering on a wavelet coefficient of the high-frequency component based on a local variance of non-noise information in the high-frequency component, to obtain the denoised high-frequency wavelet coefficients corresponding to the high-frequency components in the plurality of directions.

8 . The method according to claim 7 , wherein the estimating the local variances of non-noise information in the high-frequency components in the plurality of directions comprises:
for each of the high-frequency components in each of the plurality of directions, using a plurality of windows of different sizes to perform filtering based on the high-frequency component to obtain a filtering result corresponding to the high-frequency component in each of the windows; determining an initial local variance corresponding to the non-noise information in the high-frequency component in each of the windows based on a difference between a filtering result and a preset noise variance that correspond to a same window; and selecting a final local variance of the non-noise information in the high-frequency components from the initial local variances.

9 . The method according to claim 1 , wherein the analyzing the distribution of the fixed pattern noise in the fixed pattern noise feature map comprises:
inputting the fixed pattern noise feature map into a pre-trained authenticity detection model; analyzing, with the authenticity detection model, the distribution of fixed pattern noise in the fixed pattern noise feature map; and detecting, based on the distribution, the authenticity of the image to obtain an authenticity detection result of the image.

10 . The method according to claim 1 , wherein the detecting the authenticity of the image to obtain an authenticity detection result of the image comprises:
determining the image as an authentic image in response to the distribution matching fixed pattern noise distribution of an authentic image.

11 . An image authenticity detection apparatus, comprising:
a memory operable to store computer-readable instructions; and a processor circuitry operable to read the computer-readable instructions, the processor circuitry when executing the computer-readable instructions is configured to:
obtain an image;
remove low-frequency information from the image to obtain first image information of the image;
denoise the first image information to obtain second image information;
determine, based on a difference between the first image information and the second image information, a fixed pattern noise feature map corresponding to the image;
analyze distribution of fixed pattern noise in the fixed pattern noise feature map, the fixed pattern noise being inherent noise from a camera sensor and not interfered by image content; and
detect, b

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Han Zhou
• Zhang Zhuang
• Dong Zhiqiang

Classifications

• G06T5/70

  Denoising; Smoothing · CPC title

• G06F21/30

  Authentication, i.e. establishing the identity or authorisation of security principals · CPC title

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title

• G06V40/40

  Spoof detection, e.g. liveness detection · CPC title

• G06F21/64

  Protecting data integrity, e.g. using checksums, certificates or signatures · CPC title