IP Library › Granted Patent US 12,579,782
Granted Patent B2
US 12,579,782 · App. 18/128,708 · Granted Mar 17, 2026

Method for image processing

Inventors: Jinpeng Liu (Shanghai, CN); Tianxiang Chen (Shanghai, CN); Zijia Wang (Weifang, CN); Jiacheng Ni (Shanghai, CN); Zhen Jia (Shanghai, CN)
Assignee: Dell Products L.P.
G06V10/764G06V10/25G06V10/761G06V10/82
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,579,782
App. No.
18/128,708
Granted
Mar 17, 2026
Kind
B2
Abstract

Embodiments of the present disclosure provide a method for image processing. The method comprises extracting a target region in an image, and determining hidden vectors of the image based on the target region. The method further comprises generating a first classification result for the image based on the hidden vectors, obtaining a second classification result for the image from an image classification model, and determining trustworthiness of the image based on the first classification result and the second classification result. By using the method, a more flexible framework suitable for different applications can be provided. The framework not only can support machine learning models based on different algorithms, but also can specify data distribution more accurately. In addition, a method for independently detecting adversarial attacks is provided, thus improving the security of image detection.

Claims (66)

1 . An image processing method, comprising:

extracting a target region in an image;

determining hidden vectors of the image based on the target region;

generating a first classification result for the image based on the hidden vectors;

obtaining a second classification result for the image from an image classification model; and

determining trustworthiness of the image based on the first classification result and the second classification result;

wherein generating the first classification result for the image based on the hidden vectors comprises:

determining label categories to which the hidden vectors belong based on at least a plurality of reference truth labels; and

classifying the hidden vectors based on the label categories to obtain the first classification result.

2 . The method according to claim 1 , wherein extracting the target region in the image comprises:

determining a range of the target region to be extracted based on a type of an image processing model.

3 . The method according to claim 2 , wherein determining the range of the target region to be extracted based on the type of the image processing model further comprises:

extracting a complete target region in response to the type of the image processing model being based on a deep learning model.

4 . The method according to claim 2 , wherein determining the range of the target region to be extracted based on the type of the image processing model further comprises:

extracting part of the target region in response to the type of image processing model being based on another type different from a deep learning model.

5 . The method according to claim 1 , wherein the label categories to which the hidden vectors belong are determined based on the plurality of reference truth labels and prior knowledge.

6 . The method according to claim 5 , wherein the prior knowledge is obtained by assigning different weights to samples in a training sample set.

7 . The method according to claim 1 , wherein the plurality of reference truth labels for a training sample set are obtained by capturing and marking feature representations of training samples with the training sample set.

8 . The method according to claim 1 , wherein determining the trustworthiness of the image based on the first classification result and the second classification result comprises:

comparing the first classification result with the second classification result;

determining that the image is an untrusted image in response to the first classification result being inconsistent with the second classification result; and

determining that the image is a trusted image in response to the first classification result being consistent with the second classification result;

wherein comparing the first classification result with the second classification result comprises:

computing matching degree scores of the first classification result and the second classification result with reference truth labels, respectively; and

judging whether the matching degree scores reach a matching degree threshold defined by a user.

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

applying the hidden vectors as inputs to an image processing model to further train the image processing model for identifying the target region.

10 . A device for image processing, comprising:

at least one processor; and

memory coupled to the at least one processor and storing instructions, wherein the instructions, when executed by the at least one processor, cause the device to perform actions comprising:

extracting a target region in an image;

determining hidden vectors of the image based on the target region;

generating a first classification result for the image based on the hidden vectors;

obtaining a second classification result for the image from an image classification model; and

determining trustworthiness of the image based on the first classification result and the second classification result;

wherein generating the first classification result for the image based on the hidden vectors comprises:

determining label categories to which the hidden vectors belong based on at least a plurality of reference truth labels; and

classifying the hidden vectors based on the label categories to obtain the first classification result.

11 . The device according to claim 10 , wherein extracting the target region in the image comprises:

determining a range of the target region to be extracted based on a type of an image processing model.

12 . The device according to claim 11 , wherein determining the range of the target region to be extracted based on the type of the image processing model further comprises:

extracting a complete target region in response to the type of the image processing model being based on a deep learning model.

13 . The device according to claim 11 , wherein determining the range of the target region to be extracted based on the type of the image processing model further comprises:

extracting part of the target region in response to the type of image processing model being based on another type different from a deep learning model.

14 . The device according to claim 10 , wherein the label categories to which the hidden vectors belong are determined based on the plurality of reference truth labels and prior knowledge.

15 . The device according to claim 14 , wherein the prior knowledge is obtained by assigning different weights to samples in a training sample set.

16 . The device according to claim 10 , wherein the plurality of reference truth labels for a training sample set are obtained by capturing and marking feature representations of training samples with the training sample set.

17 . The device according to claim 10 , wherein determining the trustworthiness of the image based on the first classification result and the second classification result comprises:

comparing the first classification result with the second classification result;

determining that the image is an untrusted image in response to the first classification result being inconsistent with the second classification result; and

determining that the image is a trusted image in response to the first classification result being consistent with the second classification result;

wherein comparing the first classification result with the second classification result comprises:

computing matching degree scores of the first classification result and the second classification result with reference truth labels, respectively; and

judging whether the matching degree scores reach a matching degree threshold defined by a user.

18 . The device according to claim 10 , further comprising:

applying the hidden vectors as inputs to an image processing model to further train the image processing model for identifying the target region.

19 . A computer program product, tangibly stored in a non-transitory computer-readable storage medium and comprising machine-executable instructions, wherein the machine-executable instructions, when executed by a machine, cause the machine to perform actions comprising:

extracting a target region in an image;

determining hidden vectors of the image based on the target region;

generating a first classification result for the image based on the hidden vectors;

obtaining a second classification result for the image from an image classification model; and

determining trustworthiness of the image based on the first classification result and the second classification result;

wherein generating the first classification result for the image based on the hidden vectors comprises:

determining label categories to which the hidden vectors belong based on at least a plurality of reference truth labels; and

classifying the hidden vectors based on the label categories to obtain the first classification result.

20 . The computer program product according to claim 19 , wherein the label categories to which the hidden vectors belong are determined based on the plurality of reference truth labels and prior knowledge.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Mar 30, 2023
From: LIU, JINPENG; CHEN, TIANXIANG; WANG, ZIJIA; NI, JIACHENG; JIA, ZHEN
To: DELL PRODUCTS L.P.
Reel/Frame 063176/0365 →
Priority Claims (1)
CN 202310183393.3 · Feb 28, 2023 · national
Continuity (1)
Related Publication 20240290071A1 · Aug 29, 2024
References Cited (17)
US 11599754B2 · Soryal · 2023 [cited by examiner]
US 11762998B2 · Kuta · 2023 [cited by examiner]
US 20190238568A1 · Goswami · 2019 [cited by examiner]
US 20210150273A1 · Mustafi · 2021 [cited by examiner]
Grosse, et al. (Adversarial perturbations against Deep Neural Networks for Malware classification) (Year: 2016). [cited by examiner]
Khrulkov, et al. (Art of singular vectors and universal adversarial perturbations). (Year: 2018). [cited by examiner]
Y. Li et al., “A Review of Adversarial Attack and Defense for Classification Methods,” The American Statistician, Jan. 4, 2022, 17 pages. [cited by applicant]
D. Kalaria et al., “Detecting Adversaries, yet Faltering to Noise? Leveraging Conditional Variational AutoEncoders for Adversary Detection in the Presence of Noisy Images,” arXiv:2111.15518v2, Dec. 9, 2021, 12 pages. [cited by applicant]
U. Hwang et al., “PuVAE: A Variational Autoencoder to Purify Adversarial Examples,” arXiv:1903.00585v1, Mar. 2, 2019, 7 pages. [cited by applicant]
N. Papernot et al., “Deep k-Nearest Neighbors: Towards Confident, Interpretable and Robust Deep Learning,” arXiv:1803.04765v1, Mar. 13, 2018, 18 pages. [cited by applicant]
N. Dilokthanakul et al., “Deep Unsupervised Clustering with Gaussian Mixture Variational Autoencoders,” arXiv:1611.02648v2, Jan. 13, 2017, 12 pages. [cited by applicant]
J. Raghuram et al., “A General Framework For Detecting Anomalous Inputs to DNN Classifiers,” arXiv:2007.15147v3, Jun. 17, 2021, 27 pages. [cited by applicant]
N. Papernot et al., “Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks,” IEEE Symposium on Security & Privacy, arXiv:1511.04508v2, Mar. 14, 2016, 16 pages. [cited by applicant]
N. Carlini et al., “Defensive Distillation is Not Robust to Adversarial Examples,” arXiv:1607.04311v1, Jul. 14, 2016, 3 pages. [cited by applicant]
T. Wang et al., “Dataset Distillation,” arXiv:1811.10959v3, Feb. 24, 2020, 14 pages. [cited by applicant]
U.S. Appl. No. 17/954,536 filed in the name of Jinpeng Liu et al. on Sep. 28, 2022, and entitled “Detection of Adversarial Example Input to Machine Learning Models.” [cited by applicant]
U.S. Appl. No. 18/105,924 filed in the name of Jinpeng Liu et al. on Feb. 6, 2023, and entitled “Method, Device, and Computer Program Product for Verifying Classification Result.” [cited by applicant]