WO2024229889A1 - 人脸美丽预测方法、装置、设备及介质 - Google Patents

人脸美丽预测方法、装置、设备及介质 Download PDF

Info

Publication number
WO2024229889A1
WO2024229889A1 PCT/CN2023/095555 CN2023095555W WO2024229889A1 WO 2024229889 A1 WO2024229889 A1 WO 2024229889A1 CN 2023095555 W CN2023095555 W CN 2023095555W WO 2024229889 A1 WO2024229889 A1 WO 2024229889A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
data
training set
classifier
label
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/095555
Other languages
English (en)
French (fr)
Inventor
甘俊英
刘建强
黎慧聪
熊俊玲
谢小山
罗恒
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuyi University Fujian
Original Assignee
Wuyi University Fujian
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuyi University Fujian filed Critical Wuyi University Fujian
Priority to US18/702,055 priority Critical patent/US12620208B2/en
Publication of WO2024229889A1 publication Critical patent/WO2024229889A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06V10/7747—Organisation of the process, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G06V40/171—Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification

Definitions

  • the embodiments of the present application relate to but are not limited to the field of image recognition, and in particular to methods, devices, equipment and media for predicting the beauty of human faces.
  • the embodiments of the present application aim to solve at least one of the technical problems existing in the prior art.
  • the embodiments of the present application provide a method, apparatus, device and medium for predicting facial beauty, which can reduce the model's dependence on noise labels and enhance the utilization of unlabeled data.
  • An embodiment of the first aspect of the present invention a method for predicting beauty of a face, comprises:
  • the first image and the reweighted second image constitute a target training set, and the target training set is used to train the first model to obtain a second model;
  • the face beauty prediction model with the target classifier is used to perform face beauty prediction on the image to be predicted, and a face beauty prediction result is obtained.
  • the training set of facial beauty prediction images is classified to obtain A first image with noisy labels and a second image with non-noisy labels, comprising:
  • the training set is classified according to the joint distribution to obtain a first image with a noise label and a second image with a non-noise label.
  • the classifying the training set according to the joint distribution to obtain a first image with a noise label and a second image with a non-noise label comprises:
  • the true distribution of the training set the weight of the face beauty prediction image and the loss function, the expected risk of classification is obtained;
  • the training set is classified according to the expected risk to obtain a first image with a noise label and a second image with a non-noise label.
  • the weight of the facial beauty prediction image is determined by the joint distribution and noise rate of a training set.
  • the noise rate is a minimum value of the joint distribution of the training set within a preset range.
  • the weight when the value of the joint distribution of the training set is not equal to 0, the weight is non-negative; when the value of the joint distribution of the training set is equal to 0, the weight is equal to 0.
  • the step of training the classifier according to the first data, the second data, and the third data to obtain a target classifier comprises:
  • the classifier is trained according to the first data, the second data and the third data until the classifier converges, and the converged classifier is used as a target classifier.
  • a facial beauty prediction device comprises:
  • a reweighting unit wherein the reweighting unit is configured to: classify a training set of facial beauty prediction images to obtain a first image with a noise label and a second image with a non-noise label, reweight the second image, form a target training set with the first image and the reweighted second image, and train the first model using the target training set to obtain a second model;
  • the self-training unit is configured to: use the second model to label the image data to obtain first data with labels and second data without labels, use the second data to train a classifier to generate third data with pseudo labels, and train the classifier according to the first data, the second data and the third data to obtain to the target classifier;
  • a prediction unit is configured to: perform face beauty prediction on the image to be predicted by using a face beauty prediction model having the target classifier to obtain a face beauty prediction result.
  • An embodiment of the third aspect of the present invention is an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for predicting facial beauty when executing the computer program.
  • An embodiment of the fourth aspect of the present invention is a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the face beauty prediction method as described above.
  • the above scheme has at least the following beneficial effects: it can use noise label reweighting learning to reweight the training set, reduce the model's dependence on noise labels, and improve the model's classification performance and generalization ability; in the process of model self-training, samples with uncertain classification are regarded as noise labels, from which they are filtered or reweighted, and only samples with high confidence are retained for model training, thereby enhancing the utilization effect of unlabeled data, adaptively improving the accuracy of the model, and avoiding overfitting problems; in addition, when using unlabeled data for prediction, the prediction results with high confidence are used as new labeled data and added to the training data, thereby improving the utilization rate of samples and improving the performance of the model.
  • the algorithm can automatically adjust the model to ensure the stability and accuracy of the algorithm.
  • the algorithm is based on a simple but efficient algorithm framework and runs relatively fast.
  • FIG1 is a step diagram of a method for predicting facial beauty provided by an embodiment of the present application.
  • FIG2 is a sub-step diagram of step S100
  • FIG3 is a sub-step diagram of step S130.
  • FIG4 is a structural diagram of a facial beauty prediction device provided in an embodiment of the present application.
  • the embodiments of the present application provide a method for predicting the beauty of a face.
  • the face beauty prediction method includes:
  • Step S100 classifying a training set of facial beauty prediction images to obtain a first image with a noise label and a second image with a non-noise label;
  • Step S200 performing re-weighting processing on the second image
  • Step S300 the first image and the second image after reweighting are used to form a target training set, and the first model is trained using the target training set to obtain a second model;
  • Step S400 using the second model to perform labeling processing on the image data to obtain labeled first data and unlabeled second data;
  • Step S500 training the second data using a classifier to generate third data with a pseudo label
  • Step S600 training a classifier according to the first data, the second data and the third data to obtain a target classifier
  • Step S700 performing face beauty prediction on the image to be predicted by using a face beauty prediction model with a target classifier to obtain a face beauty prediction result.
  • LSAFBD Large Scale Asian Facial Beauty Database
  • the training set of facial beauty prediction images is classified to obtain a first image with a noise label and a second image with a non-noise label, which specifically includes but is not limited to the following steps:
  • Step S110 performing probability calculation based on the training set of the facial beauty prediction image to obtain a probability value indicating whether the facial beauty prediction image has a noise label;
  • Step S120 obtaining the joint distribution of the training set according to the probability value
  • Step S130 classifying the training set according to the joint distribution to obtain a first image with a noise label and a second image with a non-noise label.
  • the joint distribution function is also called the multidimensional distribution function.
  • the training set is classified according to the joint distribution to obtain a first image with a noise label and a second image with a non-noise label, including:
  • Step S131 obtaining the expected risk of classification according to the joint distribution of the training set, the true distribution of the training set, the weight of the face beauty prediction image and the loss function;
  • Step S132 classify the training set according to the expected risk to obtain a first image with a noise label and a second image with a non-noise label.
  • the expected risk of classification is obtained according to the joint distribution of the training set, the true distribution of the training set, the weight of the face beauty prediction image and the loss function.
  • noise label data classification method first considers that the expected risk of classification should satisfy the following relationship:
  • P D (X, Y) represents the true distribution of the training set
  • the weight of the face beauty prediction image is determined by the joint distribution and noise rate of the training set.
  • the noise rate is the minimum value of the joint distribution of the training set within a preset range.
  • the noise rate can be expressed by the following formula:
  • the weight When the value of the joint distribution of the training set is not equal to 0, the weight is non-negative; when the value of the joint distribution of the training set is equal to 0, the weight is equal to 0.
  • step S200 a re-weighting process is performed on the second image with a non-noise label.
  • reweighting can be to increase the weight of the second image with non-noise label/the first image with noisy label, reduce the weight of the second image with non-noise label/the first image with noisy label, or filter the second image with non-noise label/the first image with noisy label; the weight of the label data is adjusted through reweighting processing.
  • Reweighting refers to reweighting the training data with noisy labels so that the correct data gets more weight and the accuracy and robustness of the model are improved.
  • noisy labels are usually generated by manual or automatic annotation and have a negative impact on model performance.
  • Specific strategies should be selected according to the data set and model requirements, such as probabilistic modeling methods such as deep learning, which can effectively reduce the impact of noise labels and improve model accuracy.
  • the first image with the noise label is filtered.
  • the first image and the second image after reweighting processing constitute a target training set, and the first model is trained by using the target training set to obtain the second model.
  • the model can accurately discover the first image with noisy labels and accurately estimate the joint distribution of the first image with noisy labels and the second image with non-noisy labels.
  • step S400 the image data is labeled using the second model to obtain labeled first data and unlabeled second data.
  • the second data is used to train a classifier to generate third data with pseudo labels, wherein the classifier is obtained based on the first data with labels.
  • step S600 training a classifier according to the first data, the second data, and the third data to obtain a target classifier includes but is not limited to the following steps:
  • a classifier is trained according to the first data, the second data, and the third data until the classifier converges, and the converged classifier is used as a target classifier.
  • the model is trained and optimized by using unlabeled data.
  • the basic idea is to first use the labeled data to train the basic model, then use the model to predict the unlabeled data and select the part with high confidence as the new labeled data, and repeat the training process until the model converges or reaches the preset number of times.
  • the advantage of this algorithm is that it can make full use of unlabeled data and improve the generalization performance of the model, which is particularly suitable for applications with less labeled data.
  • step S700 face beauty prediction is performed on the image to be predicted using a face beauty prediction model with a target classifier to obtain a face beauty prediction result.
  • the above-mentioned face beauty prediction method can use noise label reweighted learning to reweight the training set, reduce the model's dependence on noise labels, and improve the model's classification performance and generalization ability; in the process of model self-training, samples with uncertain classification are regarded as noise labels, from which they are filtered or label reweighted, and only high-confidence samples are retained for model training, thereby enhancing the utilization of unlabeled data, adaptively improving the accuracy of the model, and avoiding overfitting problems.
  • the prediction results with high confidence are used as new labeled data and added to the training data, thereby improving the utilization rate of samples and the performance of the model.
  • labeled data is continuously added, and the quality of data annotation is gradually improved, making the model very robust.
  • the algorithm can automatically adjust the model to ensure the stability and accuracy of the algorithm.
  • the algorithm is based on a simple but efficient algorithm framework and runs relatively fast.
  • An embodiment of the present invention provides a facial beauty prediction device.
  • the facial beauty prediction device includes: a re-weighting unit 10 , a self-training unit 20 and a prediction unit 30 .
  • the reweighting unit 10 is configured to: classify the training set of face beauty prediction images to obtain a first image with a noise label and a second image with a non-noise label, reweight the second image, form a target training set with the first image and the reweighted second image, and use the target training set to train the first model to obtain the second model.
  • the self-training unit 20 is configured to: use the second model to label the image data to obtain labeled first data and unlabeled second data, use the second data to train the classifier to generate third data with pseudo labels, and train the classifier according to the first data, the second data and the third data to obtain a target classifier.
  • the prediction unit 30 is configured to: perform face beauty prediction on the image to be predicted by using a face beauty prediction model with a target classifier to obtain a face beauty prediction result.
  • the face beauty prediction device in this embodiment adopts the above-mentioned face beauty prediction method, and each unit of the face beauty prediction device in this embodiment corresponds to the steps of the above-mentioned face beauty prediction method, solves the same technical problems as the above-mentioned face beauty prediction method, and has the same beneficial effects as the above-mentioned face beauty prediction method.
  • An embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned method for predicting facial beauty when executing the computer program.
  • the electronic device may be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.
  • the processor can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., to execute related programs to realize the technical solutions provided in the embodiments of the present application.
  • a general-purpose CPU Central Processing Unit
  • ASIC application-specific integrated circuit
  • the memory can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM).
  • the memory can store an operating system and other application programs.
  • the relevant program codes are stored in the memory, and the processor is called to execute the methods of the embodiments of this application.
  • the input/output interface is used to realize information input and output.
  • the communication interface is used to realize the communication interaction between this device and other devices. Communication can be achieved through wired methods (such as USB, network cable, etc.) or wireless methods (such as mobile network, WIFI, Bluetooth, etc.).
  • wired methods such as USB, network cable, etc.
  • wireless methods such as mobile network, WIFI, Bluetooth, etc.
  • the bus transmits information between the various components of the device (such as the processor, memory, input/output interface, and communication interface).
  • the processor, memory, input/output interface and communication interface are connected to each other through the bus inside the device.
  • An embodiment of the present invention provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the above-mentioned face beauty prediction method.
  • the method steps in the embodiments of the present invention can be implemented or implemented by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer readable memory.
  • the method can use standard programming techniques.
  • Each program can be implemented in a high-level process or object-oriented programming language to communicate with a computer system.
  • the program can be implemented in an assembly or machine language.
  • the language can be a compiled or interpreted language.
  • the program can be run on a programmed ASIC for this purpose.
  • the processes described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context.
  • the processes described herein may be performed under the control of one or more computer systems configured with executable instructions, and may be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) that is executed collectively on one or more processors, by hardware, or a combination thereof.
  • the computer program includes a plurality of instructions that may be executed by one or more processors.
  • the method can be implemented in any type of computing platform that is operably connected to a suitable computer, including but not limited to a personal computer, a smart phone, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, etc.
  • a suitable computer including but not limited to a personal computer, a smart phone, a mainframe, a workstation, a network or distributed computing environment, a separate or integrated computer platform, or in communication with a charged particle tool or other imaging device, etc.
  • Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, an optical read and/or write storage medium, RAM, ROM, etc., so that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the process described herein.
  • machine-readable code can be transmitted via a wired or wireless network.
  • machine-readable code can be transmitted via a wired or wireless network.
  • the invention described herein includes these and other different types of non-transitory computer-readable storage media.
  • the present invention also includes the computer itself.
  • the computer program can be applied to input data to perform the functions described herein, thereby converting the input data to generate output data stored in a non-volatile memory.
  • the output information can also be applied to one or more output devices such as a display.
  • the converted data represents physical and tangible objects, including specific visual depictions of physical and tangible objects produced on the display.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

本申请实施例提供了人脸美丽预测方法、装置、设备及介质，其中方法包括对训练集进行分类出具有噪声标签的第一图像和具有非噪声标签的第二图像；对第二图像进行重加权处理；利用目标训练集对第一模型进行训练得到第二模型；利用第二模型标记图像数据得到有标签的第一数据和无标签的第二数据；通过分类器利用第二数据生成有伪标签的第三数据；根据第一数据、第二数据和第三数据训练分类器；通过具有目标分类器的模型处理待预测图像得到人脸美丽预测结果；能够减弱模型对噪声标签的依赖，提高模型的分类性能和泛化能力；增强对未标注数据的利用效果，自适应地提高模型的精度，避免过拟合问题。

Description

人脸美丽预测方法、装置、设备及介质 技术领域
本申请实施例涉及但不限于图像识别领域，尤其涉及人脸美丽预测方法、装置、设备及介质。
背景技术
目前的人脸美丽预测方法在进行模型训练时通常需要大量标签数据。同时，在标注噪声样本时，标签质量会受到人工或机器数据标注的主观因素、工具技术等因素的影响，也会带来标签噪声。标签噪声问题会大大地影响模型的精度，降低人脸美丽预测效果。
发明内容
以下是对本文详细描述的主题的概述。本概述并非是为了限制权利要求的保护范围。
本申请实施例旨在至少解决现有技术中存在的技术问题之一，本申请实施例提供了人脸美丽预测方法、装置、设备及介质，能够减弱模型对噪声标签的依赖，增强对未标注数据的利用效果。
本发明的第一方面的实施例，人脸美丽预测方法，包括：
对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像；
对所述第二图像进行重加权处理；
将所述第一图像和经重加权处理的第二图像构成目标训练集，利用所述目标训练集对第一模型进行训练得到第二模型；
利用所述第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据；
通过分类器利用所述第二数据进行训练，生成有伪标签的第三数据；
根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得到目标分类器；
通过具有所述目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
本发明的第一方面的某些实施例，所述对人脸美丽预测图像的训练集进行分类，得到具 有噪声标签的第一图像和具有非噪声标签的第二图像，包括：
根据所述人脸美丽预测图像的训练集进行概率计算，得到用于表示人脸美丽预测图像是否带有噪声标签的概率值；
根据所述概率值得到训练集的联合分布；
根据所述联合分布对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
本发明的第一方面的某些实施例，所述根据所述联合分布对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，包括：
根据训练集的联合分布、训练集的真实分布、人脸美丽预测图像的权重和损失函数，得到分类的期望风险；
根据所述期望风险对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
本发明的第一方面的某些实施例，所述人脸美丽预测图像的权重由训练集的联合分布和噪声率决定。
本发明的第一方面的某些实施例，所述噪声率为所述训练集的联合分布在预设范围内的最小值。
本发明的第一方面的某些实施例，当所述训练集的联合分布的值不等于0，所述权重非负；当所述训练集的联合分布的值等于0，所述权重等于0。
本发明的第一方面的某些实施例，所述根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得到目标分类器，包括：
根据所述第一数据、所述第二数据和所述第三数据训练所述分类器，直至所述分类器收敛，将收敛的分类器作为目标分类器。
本发明的第二方面的实施例，人脸美丽预测装置，包括：
重加权单元，所述重加权单元被配置为：对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，对所述第二图像进行重加权处理，将所述第一图像和经重加权处理的第二图像构成目标训练集，利用所述目标训练集对第一模型进行训练得到第二模型；
自训练单元，所述自训练单元被配置为：利用所述第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据，通过分类器利用所述第二数据进行训练以生成有伪标签的第三数据，根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得 到目标分类器；
预测单元，所述预测单元被配置为：通过具有所述目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
本发明的第三方面的实施例，一种电子设备，包括：存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上所述的人脸美丽预测方法。
本发明的第四方面的实施例，一种计算机可读存储介质，存储有计算机可执行指令，所述计算机可执行指令用于执行如上所述的人脸美丽预测方法。
上述方案至少具有以下的有益效果：能够利用噪声标签重加权学习来对训练集进行重加权处理，减弱模型对噪声标签的依赖，提高模型的分类性能和泛化能力；在模型自训练的过程中，将分类不确定的样本看作噪声标签，从中过滤或进行标签重加权处理，只保留高置信度的样本用于模型训练，增强对未标注数据的利用效果，自适应地提高模型的精度，避免过拟合问题；另外，在运用未标注数据进行预测时，将置信度高的预测结果作为新增的有标注数据，并加入到训练数据中，从而提高了样本的利用率，提升了模型的性能。在迭代过程中，不断增加有标注数据，并且数据标注质量逐步提高，使得模型具有很强的鲁棒性；并且，对于数据中的标签异常或噪声等因素，该算法能够自动调整模型，保证算法的稳定性和准确性。同时算法基于简单但高效的算法框架，运行速度相对较快。
附图说明
附图用来提供对本申请技术方案的进一步理解，并且构成说明书的一部分，与本申请的实施例一起用于解释本申请的技术方案，并不构成对本申请技术方案的限制。
图1是本申请的实施例所提供的人脸美丽预测方法的步骤图；
图2是步骤S100的子步骤图；
图3是步骤S130的子步骤图；
图4是本申请的实施例所提供的人脸美丽预测装置的结构图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅用以解释本申请，并不用于限定本申请。
需要说明的是，虽然在装置示意图中进行了功能模块划分，在流程图中示出了逻辑顺序，但是在某些情况下，可以以不同于装置中的模块划分，或流程图中的顺序执行所示出或描述的步骤。说明书、权利要求书或上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
下面结合附图，对本申请实施例作进一步阐述。
本申请的实施例，提供了人脸美丽预测方法。
参照图1，人脸美丽预测方法包括：
步骤S100，对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像；
步骤S200，对第二图像进行重加权处理；
步骤S300，将第一图像和经重加权处理的第二图像构成目标训练集，利用目标训练集对第一模型进行训练得到第二模型；
步骤S400，利用第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据；
步骤S500，通过分类器利用第二数据进行训练，生成有伪标签的第三数据；
步骤S600，根据第一数据、第二数据和第三数据训练分类器得到目标分类器；
步骤S700，通过具有目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
首先，获取训练集，训练集包含多张人脸图像，用人脸图像作为人脸美丽预测图像。
例如可以使用一种大规模亚洲人脸美丽数据库(Large Scale Asian Facial Beauty Database，LSAFBD)来获取人脸美丽预测图像的训练集。
参照图2，对于步骤S100，对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，具体包括但不限于以下步骤：
步骤S110，根据人脸美丽预测图像的训练集进行概率计算，得到用于表示人脸美丽预测图像是否带有噪声标签的概率值；
步骤S120，根据概率值得到训练集的联合分布；
步骤S130，根据联合分布对训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
可以理解的是，联合分布函数(joint distribution function)亦称多维分布函数。以二维情形为例，设(X，Y)是二维随机变量，x，y是任意实数，二元函数：F(x，y)＝P({X≤x ∩Y≤y})＝P(X≤x，Y≤y)，被称二维随机变量(X，Y)的分布函数，或称为X和Y的联合分布函数。
参照图3，对于步骤S130，根据联合分布对训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，包括：
步骤S131，根据训练集的联合分布、训练集的真实分布、人脸美丽预测图像的权重和损失函数，得到分类的期望风险；
步骤S132，根据期望风险对训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
对于步骤S131，根据训练集的联合分布、训练集的真实分布、人脸美丽预测图像的权重和损失函数，得到分类的期望风险。
进一步，噪声标签数据分类方法，首先考虑分类的期望风险应满足以下关系式：
其中，l表示损失函数，PD(X,Y)表示训练集的真实分布，表示训练集的联合分布，表示人脸美丽预测图像的权重。
具体地，人脸美丽预测图像的权重由训练集的联合分布和噪声率决定。
其中，为人脸美丽预测图像的权重。
噪声率为训练集的联合分布在预设范围内的最小值。噪声率可以通过以下式子表示：
当训练集的联合分布的值不等于0，权重非负；当训练集的联合分布的值等于0，权重等于0。
对于步骤S200，对具有非噪声标签的第二图像进行重加权处理。
一般地，重加权可以是增加具有非噪声标签的第二图像/具有噪声标签的第一图像的权重、降低具有非噪声标签的第二图像/具有噪声标签的第一图像的权重、或者采用对具有非噪声标签的第二图像/具有噪声标签的第一图像过滤的方式；通过重加权处理调整标签数据的权重。
重加权是指对存在噪声标签的训练数据进行重新加权，使正确的数据得到更多权重，提高模型的精度和鲁棒性。噪声标签通常由人工或自动标注产生，对模型性能影响负面。需要 对标注数据进行检测和重加权处理，缩小对模型影响。具体策略应根据数据集和模型需求选择，如基于概率建模的方法如深度学习等，有效减小噪声标签的影响，提高模型精度。
另外，对具有噪声标签的第一图像进行过滤处理。
对于步骤S300，将第一图像和经重加权处理的第二图像构成目标训练集，利用目标训练集对第一模型进行训练得到第二模型。
该模型可以准确地发现具有噪声标签的第一图像，并准确地估计具有噪声标签的第一图像和具有非噪声标签的第二图像的联合分布。
对于步骤S400，利用第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据。
对于步骤S500，通过分类器利用第二数据进行训练，生成有伪标签的第三数据。其中，分类器是根据由有标签的第一数据得到的。
对于步骤S600，根据第一数据、第二数据和第三数据训练分类器得到目标分类器，包括但不限于以下步骤：
根据第一数据、第二数据和第三数据训练分类器，直至分类器收敛，将收敛的分类器作为目标分类器。
根据以上的半监督学习，通过利用未标注数据进行模型的训练和优化。基本思想是先使用有标注数据训练基础模型，然后使用该模型对未标注数据进行预测并选取置信度高的部分作为新的标注数据，重复训练过程直到模型收敛或达到预设次数。该算法的优势在于能够充分利用未标注数据，提升模型的泛化性能，特别适用于标注数据较少情况下的应用。
对于步骤S700，通过具有目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
上述人脸美丽预测方法能够利用噪声标签重加权学习来对训练集进行重加权处理，减弱模型对噪声标签的依赖，提高模型的分类性能和泛化能力；在模型自训练的过程中，将分类不确定的样本看作噪声标签，从中过滤或进行标签重加权处理，只保留高置信度的样本用于模型训练，增强对未标注数据的利用效果，自适应地提高模型的精度，避免过拟合问题。
另外，在运用未标注数据进行预测时，将置信度高的预测结果作为新增的有标注数据，并加入到训练数据中，从而提高了样本的利用率，提升了模型的性能。在迭代过程中，不断增加有标注数据，并且数据标注质量逐步提高，使得模型具有很强的鲁棒性。
并且，对于数据中的标签异常或噪声等因素，该算法能够自动调整模型，保证算法的稳定性和准确性。同时算法基于简单但高效的算法框架，运行速度相对较快。
本发明的实施例，提供了人脸美丽预测装置。
参照图4，人脸美丽预测装置，包括：重加权单元10、自训练单元20和预测单元30。
其中，重加权单元10被配置为：对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，对第二图像进行重加权处理，将第一图像和经重加权处理的第二图像构成目标训练集，利用目标训练集对第一模型进行训练得到第二模型。
自训练单元20被配置为：利用第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据，通过分类器利用第二数据进行训练以生成有伪标签的第三数据，根据第一数据、第二数据和第三数据训练分类器得到目标分类器。
预测单元30被配置为：通过具有目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
可以理解的是，本实施例中的人脸美丽预测装置采用了如上的人脸美丽预测方法，本实施例中的人脸美丽预测装置的各单元对应上述人脸美丽预测方法的步骤，解决了和上述人脸美丽预测方法相同的技术问题，具有和上述人脸美丽预测方法相同的有益效果。
本发明的实施例，一种电子设备，包括：存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现如上所述的人脸美丽预测方法。
该电子设备可以为包括平板电脑、车载电脑等任意智能终端。
总体而言，对于电子设备的硬件结构，处理器可以采用通用的CPU(Central Processing Unit，中央处理器)、微处理器、应用专用集成电路(Application Specific Integrated Circuit，ASIC)、或者一个或多个集成电路等方式实现，用于执行相关程序，以实现本申请实施例所提供的技术方案。
存储器可以采用只读存储器(Read Only Memory，ROM)、静态存储设备、动态存储设备或者随机存取存储器(Random Access Memory，RAM)等形式实现。存储器可以存储操作系统和其他应用程序，在通过软件或者固件来实现本说明书实施例所提供的技术方案时，相关的程序代码保存在存储器中，并由处理器来调用执行本申请实施例的方法。
输入/输出接口用于实现信息输入及输出。
通信接口用于实现本设备与其他设备的通信交互，可以通过有线方式(例如USB、网线等)实现通信，也可以通过无线方式(例如移动网络、WIFI、蓝牙等)实现通信。
总线在设备的各个组件(例如处理器、存储器、输入/输出接口和通信接口)之间传输信 息。处理器、存储器、输入/输出接口和通信接口通过总线实现彼此之间在设备内部的通信连接。
本发明的实施例，一种计算机可读存储介质，存储有计算机可执行指令，所述计算机可执行指令用于执行如上所述的人脸美丽预测方法。
应当认识到，本发明实施例中的方法步骤可以由计算机硬件、硬件和软件的组合、或者通过存储在非暂时性计算机可读存储器中的计算机指令来实现或实施。所述方法可以使用标准编程技术。每个程序可以以高级过程或面向对象的编程语言来实现以与计算机系统通信。然而，若需要，该程序可以以汇编或机器语言实现。在任何情况下，该语言可以是编译或解释的语言。此外，为此目的该程序能够在编程的专用集成电路上运行。
此外，可按任何合适的顺序来执行本文描述的过程的操作，除非本文另外指示或以其他方式明显地与上下文矛盾。本文描述的过程(或变型和/或其组合)可在配置有可执行指令的一个或多个计算机系统的控制下执行，并且可作为共同地在一个或多个处理器上执行的代码(例如，可执行指令、一个或多个计算机程序或一个或多个应用)、由硬件或其组合来实现。所述计算机程序包括可由一个或多个处理器执行的多个指令。
进一步，所述方法可以在可操作地连接至合适的任何类型的计算平台中实现，包括但不限于个人电脑、智能手机、主框架、工作站、网络或分布式计算环境、单独的或集成的计算机平台、或者与带电粒子工具或其它成像装置通信等等。本发明的各方面可以以存储在非暂时性存储介质或设备上的机器可读代码来实现，无论是可移动的还是集成至计算平台，如硬盘、光学读取和/或写入存储介质、RAM、ROM等，使得其可由可编程计算机读取，当存储介质或设备由计算机读取时可用于配置和操作计算机以执行在此所描述的过程。此外，机器可读代码，或其部分可以通过有线或无线网络传输。当此类媒体包括结合微处理器或其他数据处理器实现上文所述步骤的指令或程序时，本文所述的发明包括这些和其他不同类型的非暂时性计算机可读存储介质。当根据本发明所述的方法和技术编程时，本发明还包括计算机本身。
计算机程序能够应用于输入数据以执行本文所述的功能，从而转换输入数据以生成存储至非易失性存储器的输出数据。输出信息还可以应用于一个或多个输出设备如显示器。在本发明优选的实施例中，转换的数据表示物理和有形的对象，包括显示器上产生的物理和有形对象的特定视觉描绘。
尽管已经示出和描述了本申请的实施方式，本领域的普通技术人员可以理解：在不脱离本申请的原理和宗旨的情况下可以对这些实施方式进行多种变化、修改、替换和变型，本申 请的范围由权利要求及其等同物限定。
以上是对本申请的较佳实施进行了具体说明，但本申请并不限于实施例，熟悉本领域的技术人员在不违背本申请精神的前提下可做作出种种的等同变形或替换，这些等同的变形或替换均包含在本实施例所限定的范围内。

Claims (10)

  1. 人脸美丽预测方法，其特征在于，包括：
    对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像；
    对所述第二图像进行重加权处理；
    将所述第一图像和经重加权处理的第二图像构成目标训练集，利用所述目标训练集对第一模型进行训练得到第二模型；
    利用所述第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据；
    通过分类器利用所述第二数据进行训练，生成有伪标签的第三数据；
    根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得到目标分类器；
    通过具有所述目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
  2. 根据权利要求1所述的人脸美丽预测方法，其特征在于，所述对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，包括：
    根据所述人脸美丽预测图像的训练集进行概率计算，得到用于表示人脸美丽预测图像是否带有噪声标签的概率值；
    根据所述概率值得到训练集的联合分布；
    根据所述联合分布对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
  3. 根据权利要求1所述的人脸美丽预测方法，其特征在于，所述根据所述联合分布对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，包括：
    根据训练集的联合分布、训练集的真实分布、人脸美丽预测图像的权重和损失函数，得到分类的期望风险；
    根据所述期望风险对所述训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像。
  4. 根据权利要求3所述的人脸美丽预测方法，其特征在于，所述人脸美丽预测图像的权重由训练集的联合分布和噪声率决定。
  5. 根据权利要求4所述的人脸美丽预测方法，其特征在于，所述噪声率为所述训练集的联合分布在预设范围内的最小值。
  6. 根据权利要求4所述的人脸美丽预测方法，其特征在于，当所述训练集的联合分布的值不等于0，所述权重非负；当所述训练集的联合分布的值等于0，所述权重等于0。
  7. 根据权利要求1所述的人脸美丽预测方法，其特征在于，所述根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得到目标分类器，包括：
    根据所述第一数据、所述第二数据和所述第三数据训练所述分类器，直至所述分类器收敛，将收敛的分类器作为目标分类器。
  8. 人脸美丽预测装置，其特征在于，包括：
    重加权单元，所述重加权单元被配置为：对人脸美丽预测图像的训练集进行分类，得到具有噪声标签的第一图像和具有非噪声标签的第二图像，对所述第二图像进行重加权处理，将所述第一图像和经重加权处理的第二图像构成目标训练集，利用所述目标训练集对第一模型进行训练得到第二模型；
    自训练单元，所述自训练单元被配置为：利用所述第二模型对图像数据进行标记处理，得到有标签的第一数据和无标签的第二数据，通过分类器利用所述第二数据进行训练以生成有伪标签的第三数据，根据所述第一数据、所述第二数据和所述第三数据训练所述分类器得到目标分类器；
    预测单元，所述预测单元被配置为：通过具有所述目标分类器的人脸美丽预测模型对待预测图像进行人脸美丽预测，得到人脸美丽预测结果。
  9. 一种电子设备，包括：存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至7中任一项所述的人脸美丽预测方法。
  10. 一种计算机可读存储介质，其特征在于，存储有计算机可执行指令，所述计算机可执行指令用于执行如权利要求1至7中任一项所述的人脸美丽预测方法。
PCT/CN2023/095555 2023-05-10 2023-05-22 人脸美丽预测方法、装置、设备及介质 Ceased WO2024229889A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US18/702,055 US12620208B2 (en) 2023-05-10 2023-05-22 Facial beauty prediction method, apparatus, device and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310525485.5 2023-05-10
CN202310525485.5A CN116824660A (zh) 2023-05-10 2023-05-10 人脸美丽评分方法、装置、设备及介质

Publications (1)

Publication Number Publication Date
WO2024229889A1 true WO2024229889A1 (zh) 2024-11-14

Family

ID=88141951

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/095555 Ceased WO2024229889A1 (zh) 2023-05-10 2023-05-22 人脸美丽预测方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN116824660A (zh)
WO (1) WO2024229889A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119169682B (zh) * 2024-08-30 2025-11-11 五邑大学 人脸评分模型的训练方法、人脸评分方法及相关装置

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114445885A (zh) * 2022-01-18 2022-05-06 深圳数联天下智能科技有限公司 训练多目标检测模型的方法、检测方法及相关装置
CN114463810A (zh) * 2022-01-17 2022-05-10 北京三快在线科技有限公司 一种人脸识别模型的训练方法和装置
CN114611607A (zh) * 2022-03-14 2022-06-10 腾讯科技（上海）有限公司 一种模型训练方法以及相关装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11663486B2 (en) * 2020-06-23 2023-05-30 International Business Machines Corporation Intelligent learning system with noisy label data
CN111966823B (zh) * 2020-07-02 2022-04-22 华南理工大学 一种面向标签噪声的图节点分类方法
CN115761288A (zh) * 2022-11-03 2023-03-07 西安交通大学 基于数据自选择与标记自校正算法的图像偏差数据分类方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114463810A (zh) * 2022-01-17 2022-05-10 北京三快在线科技有限公司 一种人脸识别模型的训练方法和装置
CN114445885A (zh) * 2022-01-18 2022-05-06 深圳数联天下智能科技有限公司 训练多目标检测模型的方法、检测方法及相关装置
CN114611607A (zh) * 2022-03-14 2022-06-10 腾讯科技（上海）有限公司 一种模型训练方法以及相关装置

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
GAN JUNYING, WU BICHENG, ZHAI YIKUI, HE GUOHUI, MAI CHAOYUN, BAI ZHENFENG: "Self-correcting noise labels for facial beauty prediction", JOURNAL OF IMAGE AND GRAPHICS, ZHONGGUO TUXIANG TUXING XUEHUI, CN, vol. 27, no. 8, 1 August 2022 (2022-08-01), CN , pages 2487 - 2495, XP093237051, ISSN: 1006-8961, DOI: 10.11834/jig.210125 *

Also Published As

Publication number Publication date
US20250239058A1 (en) 2025-07-24
CN116824660A (zh) 2023-09-29

Similar Documents

Publication Publication Date Title
US12100192B2 (en) Method, apparatus, and electronic device for training place recognition model
US10878173B2 (en) Object recognition and tagging based on fusion deep learning models
US11875512B2 (en) Attributionally robust training for weakly supervised localization and segmentation
CN113574566B (zh) 目标检测网络构建优化方法、装置、设备、介质及产品
CN115082920B (zh) 深度学习模型的训练方法、图像处理方法和装置
CN119068504B (zh) 一种基于机器视觉的票据识别方法及装置
CN111985620A (zh) 一种神经网络模型知识蒸馏方法、系统、设备及介质
CN111862040B (zh) 人像图片质量评价方法、装置、设备及存储介质
CN108492301A (zh) 一种场景分割方法、终端及存储介质
WO2023025063A1 (zh) 图像信号处理器优化方法及设备
CN115471714A (zh) 数据处理方法、装置、计算设备和计算机可读存储介质
CN108154191A (zh) 文档图像的识别方法和系统
CN113361384B (zh) 人脸识别模型压缩方法、设备、介质及计算机程序产品
CN109325512A (zh) 图像分类方法及装置、电子设备、计算机程序及存储介质
WO2024229889A1 (zh) 人脸美丽预测方法、装置、设备及介质
CN113139466B (zh) 基于单隐层神经网络的图像识别方法及相关设备
CN117058498B (zh) 分割图评估模型的训练方法、分割图的评估方法及装置
CN118760844A (zh) 基于AutoEdge的智算平台自动优化方法、装置、设备及存储介质
CN108229552B (zh) 一种模型处理方法、装置及存储介质
CN118588313A (zh) 医院数据字典映射方法、装置、计算设备及存储介质
CN117058739A (zh) 一种人脸聚类更新方法及装置
US12620208B2 (en) Facial beauty prediction method, apparatus, device and storage medium
JP5652250B2 (ja) 画像処理プログラム及び画像処理装置
CN114429579A (zh) 图像处理、分类方法及装置、电子设备及存储介质
CN117236297B (zh) 一种基于pdf跨页签字的实现方法

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 18702055

Country of ref document: US

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23936202

Country of ref document: EP

Kind code of ref document: A1

WWP Wipo information: published in national office

Ref document number: 18702055

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE