IP Library › Granted Patent US 12,541,675
Granted Patent B2
US 12,541,675 · App. 18/173,591 · Granted Feb 3, 2026

Distances between distributions for the belonging-to-the-distribution measurement of the image

Inventors: Stepan Alekseevich Komkov (Moscow, RU); Aleksandr Aleksandrovich Petiushko (Moscow, RU); Ivan Leonidovich Mazurenko (Moscow, RU); Jiang Li (Shenzhen, CN)
Assignee: Huawei Technologies Co., Ltd.
G06N3/0499G06V10/82G06V40/172
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 12,541,675
App. No.
18/173,591
Granted
Feb 3, 2026
Kind
B2
Abstract

The present disclosure relates to processing input data by a neural network. Methods and apparatuses of some embodiments process the input data by at least one layer of the neural network and obtain thereby a feature tensor. Then, the distribution of the obtained feature tensors estimated. Another distribution is obtained. Such other distribution may be a distribution of another input data, or a distribution obtained by combining a plurality of distributions obtained for respective plurality of some input data. Then a distance value indicative of a distance between the two distributions is calculated and based thereon, a characteristic of the input data is determined. The characteristic may be pertinence to a certain class of data or a detection of out-of-distribution data or determination of reliability of a class determination or the like.

Claims (55)

1 . A method for processing first input data by a neural network, which is a feed-forward neural network, the method comprising:

processing the first input data by at least one layer of the neural network to obtain a first feature tensor, wherein the first input data is image data including c channels with c being an integer equal to or larger than 1;

estimating a first distribution of the obtained first feature tensor;

obtaining a second distribution;

determining a distance value indicative of a distance between the first distribution and the second distribution;

determining a characteristic of the first input data based on the determined distance value, wherein

the obtaining of the second distribution includes:

processing of second input data by the at least one layer to obtain a second feature tensor; and

estimating the second distribution of the obtained second feature tensor; and

the second distribution is obtained by averaging of a plurality of distributions determined for respective plurality of input data belonging to a same class, and

wherein determining of the characteristic of the first input data includes comparing the distance value with a threshold, and, based on the comparison, estimating reliability of the first input data belonging to said same class and wherein the set of classes are open and at least one new class is defined during processing.

2 . The method according to claim 1 , wherein the estimating of the first distribution includes:

obtaining a number, n, of value intervals of the first feature tensor values, and

determining, for each of the n value intervals, number of occurrences of values belonging to said value interval among the first feature tensor values.

3 . The method according to claim 2 , wherein the obtaining of the n value intervals includes at least one of (i) the number n based on the dimensions of the first feature tensor, and (ii) determining the value interval length based on entropy of the first feature tensor values.

4 . The method according to claim 1 , wherein, in the determining of the characteristic of the first input data, the characteristic is at least one of

a class among a plurality of predetermined classes of data, and

whether the first input data belong to one of the predetermined classes of data.

5 . The method according to claim 1 , wherein

the determining of the characteristic of the first input data comprises determining similarity of the first input data to the second input data by a similarity metric being a function of said distance value.

6 . The method according to claim 5 , wherein the similarity metric is further a function of at least one of a feature tensor obtained by processing the first input data with all layers of the neural network and a feature tensor obtained by processing the second data with all layers of the neural network.

7 . The method according to claim 5 , wherein the function is a monotonically increasing function of said distance value.

8 . The method according to claim 7 , wherein the function ψ is given by ψ(s, d)=p 1 (s)+w·p 2 (min(d, Θ)), with p i (x)=x ai , wherein ai is a positive integer, i is 1 or 2, w is a predetermined weight factor, Θ is a predetermined maximum distance.

9 . The method according to claim 1 , wherein, the determining of the characteristic of the first input data includes:

comparing the distance value with a threshold; and

based on the comparison, estimating reliability of the first input data belonging to said same class.

10 . The method according to claim 1 , wherein the distance value is calculated based on Hellinger distance.

11 . The method according to claim 10 , wherein the distance value is calculated by approximating the Hellinger distance as a sum of squared differences projected to a space reduced by applying Principal Component Analysis.

12 . The method according to claim 1 , further comprising determining of said at least one layer of the neural network as the layer of which the output feature tensor provides the maximum classification accuracy.

13 . The method according to claim 1 , wherein

the steps of processing the first input data, estimating a first distribution of the obtained first feature tensor, and determining a distance value are performed separately for each channel c; and

the determining of the characteristic of the first input data is based on an aggregation of the distance values determined for each channel.

14 . The method according to claim 1 , wherein the method is used for face recognition.

15 . A non-transitory medium storing instructions which, when executed on one or more processors, perform steps comprising:

processing the first input data by at least one layer of the neural network to obtain a first feature tensor, wherein the first input data is image data including c channels with c being an integer equal to or larger than 1;

estimating a first distribution of the obtained first feature tensor;

obtaining a second distribution;

determining a distance value indicative of a distance between the first distribution and the second distribution;

determining a characteristic of the first input data based on the determined distance value, wherein

the obtaining of the second distribution includes:

processing of second input data by the at least one layer to obtain a second feature tensor; and

estimating the second distribution of the obtained second feature tensor; and

the second distribution is obtained by averaging of a plurality of distributions determined for respective plurality of input data belonging to a same class, and

wherein determining of the characteristic of the first input data includes comparing the distance value with a threshold, and, based on the comparison, estimating reliability of the first input data belonging to said same class and wherein the set of classes are open and at least one new class is defined during processing.

16 . A signal processing apparatus for processing first input data by a neural network, which is a feed-forward neural network, the signal processing apparatus comprising processing circuitry configured to:

process the first input data by at least one layer of the neural network to obtain a first feature tensor, wherein the first input data is image data including c channels with c being an integer equal to or larger than 1;

estimate a first distribution of the obtained first feature tensor;

obtain a second distribution;

determine a distance value indicative of a distance between the first distribution and the second distribution;

determine a characteristic of the first input data based on the determined distance value, wherein

the obtain of the second distribution includes:

process of second input data by the at least one layer to obtain a second feature tensor; and

estimate the second distribution of the obtained second feature tensor; and

the second distribution is obtained by averaging of a plurality of distributions determined for respective plurality of input data belonging to a same class, and

wherein determine the characteristic of the first input data includes compare the distance value with a threshold, and, based on the comparison, estimate reliability of the first input data belonging to said same class and wherein the set of classes are open and at least one new class is defined during processing.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Sep 24, 2025
From: KOMKOV, STEPAN ALEKSEEVICH; PETIUSHKO, ALEKSANDR ALEKSANDROVICH; MAZURENKO, IVAN LEONIDOVICH; LI, JIANG
To: HUAWEI TECHNOLOGIES CO., LTD.
Reel/Frame 072365/0682 →
Continuity (2)
Continuation PCTRU2020000450 · Aug 25, 2020
Related Publication 20230229897A1 · Jul 20, 2023
References Cited (37)
US 9767381B2 · Rodríguez-Serrano et al. · 2017 [cited by applicant]
US 10346464B2 · Ye · 2019 [cited by applicant]
US 10558750B2 · Lu et al. · 2020 [cited by applicant]
US 20210374524A1 · Feng · 2021 [cited by examiner]
CN 109460777A · 2019 [cited by applicant]
CN 109472240A · 2019 [cited by applicant]
CN 109977887A · 2019 [cited by applicant]
Herrmann et al: “Low-Resolution Convolutional Neural Networks for Video Face Recognition”, IEEE, 2016 (Year: 2016). [cited by examiner]
Hsu et al: “The Effect of Distance Function for NN Classifier with Local Binary Pattern Descriptors”, 2018 (Year: 2018). [cited by examiner]
Aizenbud et al.: “PCA-Based Out-of-Sample Extension for Dimensionality Reduction”, 2015 (Year: 2015). [cited by examiner]
Liang et al., “Enhancing the Reliability of Out-of-Distribution Image Detection in Neural Networks,” Published as a conference paper at ICLR 2018, arXiv: 1706.02690v4 [cs.LG], XP081281629, total 27 pages (Feb. 25, 2018). [cited by applicant]
Grother et al., “Face Recognition Technology Evaluation (FRTE) Part 1: Verification,” NISTIR XXXX Draft, National Institute of Standards and Technology, total 939 pages (Aug. 9, 2024). [cited by applicant]
Lee et al., “A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks,” arXiv: 1807.03888v2 [stat.ML], XP081055670, total 20 pages (Oct. 27, 2018). [cited by applicant]
Dhamija et al., “Reducing Network Agnostophobia,” arXiv: 1811.04110v2 [cs.CV], total 14 pages (Dec. 23, 2018). [cited by applicant]
Quintaniha et al., “Detecting Out-of-Distribution Samples Using Low-Order Deep Feature Statistics,” Under review as a conference paper at ICLR 2019, XP055603220, total 17 pages (Nov. 19, 2018). [cited by applicant]
Ruff et al., “Deep One-Class Classification,” Proceedings of the 35th International Conference on Machine Learning, total 10 pages (2018). [cited by applicant]
Sastry et al., “Detecting Out-of-Distribution Examples with In-Distribution Examples and Gram Matrices,” arXiv: 1912.12510v2 [cs.LG], XP081575090, total 21 pages (Jan. 9, 2020). [cited by applicant]
Shu et al., “Unseen Class Discovery in Open-World Classification,” arXiv: 1801.05609v1 [cs.LG], total 10 pages (Jan. 17, 2018). [cited by applicant]
Ditzler et al., “Hellinger Distance Based Drift Detection for Nonstationary Environments,” 2011 IEEE Symposium on Computational Intelligence in Dynamic and Uncertain Environments (CIDUE), XP032003065, total 8 pages, Ins… [cited by applicant]
Yoshihashi et al., “Classification-Reconstruction Learning for Open-Set Recognition,” arXiv: 1812.04246v3 [cs.CV], total 11 pages (Oct. 6, 2019). [cited by applicant]
Oza et al., “C2AE: Class Conditioned Auto-Encoder for Open-set Recognition,” arXiv: 1904.01198v1 [cs.CV], total 13 pages (Apr. 2, 2019). [cited by applicant]
Rudd et al., “The Extreme Value Machine,” IEEE Transactions on Pattern Analysis and Machine Intelligence, arXiv: 1506.06112v4 [cs.LG], total 12 pages (May 21, 2017). [cited by applicant]
Yu et al., “Unsupervised Out-of-Distribution Detection by Maximum Classifier Discrepancy,” total 9 pages (Oct. 27-Nov. 2, 2019). [cited by applicant]
Vyas et al., “Out-of-Distribution Detection Using an Ensemble of Self Supervised Leave-out Classifiers,” total 15 pages (Sep. 8, 2018). [cited by applicant]
Hendrycks et al., “A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks,” Published as a conference paper at ICLR 2017, arXiv: 1610.02136v3 [cs.NE], total 12 pages (Oct. 3, 2018). [cited by applicant]
Liang et al., “Enhancing the Reliability of Out-Of-Distribution Image Detection in Neural Networks,” Published as a conference paper at ICLR 2018, arXiv: 1706.02690v5 [cs.LG], total 15 pages (Aug. 30, 2020). [cited by applicant]
Meller et al., “Image Classification with Feed-Forward Neural Networks,” CEUR-WS.org/VOL-2694/p4.pdf, total 7 pages (May 20, 2020). [cited by applicant]
Bendale et al., “Towards Open Set Deep Networks,” arXiv: 1511.06233v1 [cs.CV], total 14 pages (Nov. 19, 2015). [cited by applicant]
Shi et al., “Probabilistic Face Embeddings,” arXiv: 1904.09658v4 [cs.CV], total 12 pages (Aug. 7, 2019). [cited by applicant]
Gómez et al., “Understanding Categorical Cross-Entropy Loss, Binary Cross-Entropy Loss, Softmax Loss, Logistic Loss, Focal Loss and all those confusing names,” total 12 pages (May 23, 2018). [cited by applicant]
Lecun et al., “The MNIST Database of Handwritten Digits,” URL: https://yann.lecun.com/exdb/mnist/, total 8 pages (Dec. 4, 2024). [cited by applicant]
“Weibull distribution,” [online] https://en.wikipedia.org/wiki/Weibull_distribution, total 10 pages (Aug. 20, 2024). [cited by applicant]
“Understanding of Convolutional Neural Network (CNN)—Deep Learning,” total 13 pages (Oct. 18, 2023). [cited by applicant]
“Hellinger distance,” [online] https://en.wikipedia.org/wiki/Hellinger_distance, total 4 pages (Dec. 20, 2023). [cited by applicant]
Kumar, “Understanding Principal Component Analysis,” total 10 pages (Jan. 2, 2018). [cited by applicant]
Pablo Ruiz: “Understanding and visualizing ResNets”.Oct. 8, 2018, total 12 pages. [cited by applicant]
“Entropy (information theory),” [online] https://en.wikipedia.org/wiki/Entropy_(information_theory), total 15 pages (Aug. 22, 2024). [cited by applicant]