Age estimation method based on diffusion model

Abstract

The invention discloses an age estimation method based on a diffusion model, and relates to the technical field of face recognition. Comprising the following steps: constructing a face age estimation model, and inputting a data set into the face age estimation model; performing rearrangement marking and noise removal through a pairing comparison module to obtain a face age image; extracting features of the face age image through a trunk feature extractor, and decoupling the extracted features into additional variables; on the basis of the non-age features, Gaussian noise is gradually added to the face age image in multiple iterations through the forward process of a diffusion model, and a target image with noise is obtained; synthesizing the target image with noise and the age feature through a reverse process of a diffusion model to obtain an age corresponding to the face age image; and inputting a to-be-detected face image into the trained face age estimation model to obtain an age corresponding to the to-be-detected face image. The method can effectively separate and recombine age-related and irrelevant features, and improves the accuracy of face age estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of face recognition, and in particular to an age estimation method based on a diffusion model. Background Art

[0002] Facial age estimation technology has important applications in multiple fields, including security control, human-computer interaction, and biometrics. As people age, their facial features gradually change. Due to factors such as muscle relaxation, sagging skin, and bone changes, the facial contour loses its youthful appearance. The human face contains many important pieces of information related to personal characteristics, the most important of which is age. Therefore, determining age based on facial features has become a hot research topic.

[0003] Over the past few decades, facial age estimation has been extensively studied, with methods primarily categorized into three categories: classification, regression, and ranking. Classification methods treat different age labels as independent categories, while regression methods predict age labels based on their continuity. Ranking methods estimate age by learning the relative differences between ages. In recent years, age estimation methods based on label distributions have also garnered significant attention. These methods combine ordinal and variable boundaries to minimize the influence of head classes on predictions of tail samples. Early research focused on characterizing age, but these features struggle to capture discriminative information for accurate age classification or regression. Furthermore, these studies failed to consider the ambiguity of age labels that arises during aging, specifically the ordinal relationship between age and factors such as ethnicity and background. These factors can introduce significant bias in datasets and, in turn, in the learning process. Learning from data with noisy labels is an important research direction in deep learning. Some methods guide model learning by constructing noisy models or building knowledge graphs using small amounts of clean data. However, these methods rely on small amounts of cleanly labeled data or specific assumptions about the noisy models, and their effectiveness is often questioned in the face of complex noisy label scenarios.

[0004] Meta-learning methods proposed in recent years overcome the impact of noisy labels by constructing noisy models or using data augmentation and meta-learning. While these methods can balance meta-weights in facial aging datasets by learning instance-level boundaries, addressing class variance and class imbalance, and achieving unbiased age classification, they are, however, highly demanding on the quality of early training datasets and sensitive to threshold selection compared to training iterations, making it difficult to maintain balance across all datasets. This disrupts the ordered relationship between age categories, hindering effective feature decoupling in existing models and leading to model instability due to incorrectly learning age features. Summary of the Invention

[0005] Based on this, it is necessary to provide an age estimation method based on a diffusion model to address the above technical issues.

[0006] An embodiment of the present invention provides an age estimation method based on a diffusion model, comprising: The dataset is obtained by uniformly sampling from a face dataset in an unconstrained environment containing noise samples; Inputting the data set into a facial age estimation model, the facial age estimation model comprising a paired comparison module and a double-constrained diffusion model connected in sequence, the double-constrained diffusion model comprising a backbone feature extractor and a diffusion model connected in sequence; The paired comparison module performs rearrangement labeling and noise removal to obtain an ordered face age image with noise-free labels. The backbone feature extractor extracts features from the face age image and decouples the extracted features into additional variables including age features and non-age features. The forward diffusion process of the diffusion model gradually adds Gaussian noise to the ordered face age image with noise-free labels over multiple iterations to obtain a noisy target image. The reverse generation process of the diffusion model performs neural denoising on the noisy target image and age features to achieve aging reconstruction of the face age image and obtain the age corresponding to the face age image, thereby obtaining a trained facial age estimation model. The face image to be tested is input into the trained facial age estimation model to obtain the age corresponding to the face image to be tested.

[0007] Optionally, rearrange labeling and noise removal are performed through a paired comparison module to obtain an ordered face age image with noise-free labels, specifically including: Initialize the sample fraction of each sample in the dataset ,in, Indicates a score, represents a sample; Compare the samples in the dataset in pairs and score the image of the older individual in each pair; According to the cumulative scores of the samples, we continuously compare the samples with the same scores and iteratively improve the ranking of the samples to obtain the sorted data set. The formula is expressed as: ; in, Representative samples The score, Represents another sample The score, represents the sorted dataset, Represents a constant within (0, 1); The sorted dataset is first regularized, and the data samples are mapped to different age ranges according to the specific characteristics of each dataset; The mapped values ​​are rounded based on the following formula to obtain an ordered face age image with noise-free labels: ; in, represents the mean of the sorted data set, Represents the standard deviation of the sorted data set, interval Represents the lower and upper bounds of the sorted dataset.

[0008] Optionally, the method further includes testing the unsorted dataset using a deep neural network after rearrangement labeling and noise removal by the paired comparison module, and selecting samples with larger prediction errors based on the following formula: ; in, represents the label predicted by the deep neural network, Indicates the true age label; According to the comparison result between the prediction error and the set threshold, the re-labeled samples are selected, and the selected re-labeled samples are added to the sorted data set to obtain the data set of the facial age estimation model.

[0009] Optionally, Gaussian noise is gradually added to the ordered face age images with noise-free labels in multiple iterations through a forward diffusion process of a diffusion model to obtain a noisy target image, specifically including: In the forward process of the diffusion model, a forward Markov diffusion process guided by the non-age features in the additional variables is defined. The forward Markov diffusion process gradually adds Gaussian noise to the ordered face age images with noise-free labels in multiple iterations. The formula is expressed as: ; ; The intermediate steps are marginalized based on the following formula to obtain the noisy target image: ; in, represents the number of iterations, represents ordered and noise-free labeled face age images, represents a scalar hyperparameter, represents the sum of scalar hyperparameters over multiple iterations, Represents a face age image.

[0010] Optionally, the noisy target image and age features are subjected to neural denoising through the inverse generation process of the diffusion model to achieve aging reconstruction of the face age image and obtain the age corresponding to the face age image: In the reverse generation process of the diffusion model, the noisy target image and age features are used as the input of the neural denoising model, and the formula is expressed as: ; In the synthesis stage, the additional variables are guided as conditions, and the processing formula is expressed as: ; in, represents non-age characteristics in the additional variables, represents ordered and noise-free labeled face age images, represents the face age image, represents the noise vector.

[0011] Optionally, the method further includes conditioning the age feature as an input of a neural denoising model to predict the noise vector; the objective function of the conditioning process is: ; During the diffusion process, the gradient is continuously back-propagated to improve the decoupling ability of the reverse process of the diffusion model, using the mean absolute error loss: ; in, Indicates the true age label; The overall objective function formula is expressed as: ; in, is the conditional process objective function, is the mean absolute error loss.

[0012] Optionally, implementing facial aging reconstruction specifically includes two facial aging reconstruction processes: a facial aging reconstruction process involving the same age but different identities and a facial aging reconstruction process involving the same identity but different ages; The reconstruction process is expressed as: ; ; in, and Represents two face images of the same age but different identities, and Two face images representing the same identity but different ages, and Respectively represent and age characteristics, and Respectively represent and Non-age characteristics, and Respectively represent and age characteristics, and Respectively represent and non-age characteristics.

[0013] Compared with the prior art, the age estimation method based on the diffusion model provided by the embodiment of the present invention has the following beneficial effects: The present invention decouples the features of the extracted facial age image into age features and non-age features that are conditionally restricted during the diffusion model processing process, so as to separate features related to age from features not related to age. Age features and non-age features are used as constraints in the forward process and the reverse process respectively, effectively separating and recombining features related to age from features not related to age, thereby improving the accuracy of facial age estimation.

[0014] In addition, the present invention uses a paired comparison module to perform rearrangement labeling and noise removal to obtain ordered facial age images with noise-free labels, which can ensure the accuracy of sample labeling, reduce the negative impact of noise labels on performance, and thus maintain the continuity of age estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a model flow of an age estimation method based on a diffusion model provided in one embodiment; Figure 2 The active learning flowchart of an age estimation method based on a diffusion model provided in one embodiment is shown. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0017] A dual-constraint diffusion model (DCDM) is proposed to compare aging appearance for facial age estimation, addressing the critical issue of noisy labels. First, a paired comparison module is designed to reorder the dataset labels to ensure continuity and accurate labeling. Second, a dual-constraint diffusion model (DCDM) is employed to effectively separate and reassemble age-related and irrelevant features, thereby facilitating the generation of high-fidelity and continuously age-progressive facial representations. Finally, an active learning framework is used to optimize model parameters using age-difference information.

[0018] In one embodiment, a method for estimating age based on a diffusion model is provided, such as Figure 1 As shown, the method includes: The dataset is obtained by uniformly sampling from a face dataset in an unconstrained environment containing noise samples.

[0019] The dataset is input into a facial age estimation model, which includes a paired comparison module and a double-constrained diffusion model connected in sequence, and the double-constrained diffusion model includes a backbone feature extractor and a diffusion model connected in sequence.

[0020] The paired comparison module performs rearrangement labeling and noise removal, resulting in an ordered, noise-free facial age image. The backbone feature extractor extracts features from the facial age image and decouples the extracted features into additional variables, including age and non-age features. The forward diffusion process of the diffusion model gradually adds Gaussian noise to the ordered, noise-free facial age image over multiple iterations, resulting in a noisy target image. The reverse generation process of the diffusion model then performs neural denoising on the noisy target image and age features to achieve aging reconstruction of the facial age image and obtain the age corresponding to the facial age image, ultimately yielding a trained facial age estimation model.

[0021] The face image to be tested is input into the trained facial age estimation model to obtain the age corresponding to the face image to be tested.

[0022] One is the paired comparison module Paired Comparison, and the other is the double-constrained diffusion model DCDM. It requires two types of input: one is the dataset in an unconstrained environment , these data are input into Paired Comparison. Specifically, the dataset It is uniformly sampled from a face dataset in an unconstrained environment containing noise samples. Another input is a face age image ,These are clean samples with noise-free labels that are sorted by the paired comparison module.

[0023] The dual-constrained diffusion model DCDM consists of two main parts: a backbone feature extractor And a diffusion model including forward process and reverse process. For the input image , the features are extracted by the backbone ResNet-34 network, and the noise image Obtained by adding noise in the forward process of the diffusion model, and restoring the image in the reverse process To learn the noise. During the diffusion process, the age feature is defined and non-age characteristics Constraints are imposed in both the forward and reverse processes to ensure high-quality sample generation. During the generation process, gradients are continuously back-propagated to enhance the classifier's decoupling capabilities, thereby improving discriminative capabilities. This effectively mitigates the impact of noisy samples and effectively separates age and non-age features. Finally, an active learning strategy is employed to selectively label samples, thereby improving classification performance.

[0024] 1. Paired Comparison Module The paired comparison module is designed to sort the data set by the paired comparison algorithm. Assume that the input data set of the algorithm is , initialize the sample fraction ,in , Denote the scores and samples. In paired comparisons, due to the presence of noisy age labels, ranking no longer relies on labels. With age, faces exhibit increasing amounts of detail, such as wrinkles or changes in skin texture. Considering the varying facial details at different ages, paired comparisons are defined as follows. First, samples are manually compared in pairs, with the image of the older individual in each pair scored. Based on their cumulative scores, samples with the same score are continuously compared, iteratively improving their ranking.

[0025] Here, Representative samples The score, Represents another sample The score, represents the sorted dataset, is a constant in the interval (0, 1). To ensure that the sorted samples correctly reflect age labels, the sorted dataset is first normalized. Then, the data samples are mapped to different age intervals based on the specific characteristics of each dataset. Finally, these values ​​are rounded to obtain a well-structured dataset with accurate age labels.

[0026] Thus, the dataset Samples in Determined by: Here, represents the mean of the sorted data set, Represents the standard deviation of the sorted data set, interval Represents the lower and upper bounds of the sorted dataset.

[0027] 2. Dual-Constraint Diffusion Model 1. Training process: The dual-constrained diffusion model consists of two components: a backbone feature extractor and a diffusion model. Specifically, the input image is fed into two networks. One of them is a feature extractor. , which starts from the input sample Extract features. After the feature extraction structure, the output is as follows

[0028] Next, the extracted features are decoupled into and , these features will be used for facial aging reconstruction and conditional constraints through the diffusion process. In order to achieve facial aging reconstruction, and Compared with CycleGAN, which requires multiple models to complete the aging reconstruction process, the present invention only needs one model to achieve it. Two facial aging reconstruction processes are designed: one involves the same age but different identities, and the other involves and Another involves the same identity but different age, represented by and The exchange and reconstruction process is as follows:

[0029] in, and Respectively represent and age characteristics, and Respectively represent and In this process, and of Despite the swap reconstruction, they are still able to reconstruct their appearance because they have the same age characteristics. and of By exchanging and reconstructing, they can also reconstruct their appearance because they have the same identity characteristics.

[0030] Based on the above facial aging reconstruction process, a double-constrained diffusion model was designed, which was subject to additional variables (age characteristics and non-age characteristics ) conditional restrictions. In the forward process, compared with the traditional diffusion model, we first define a Forward Markov Diffusion Process as Conditional Guidance , the process is In the iterations, we gradually move towards a high-resolution image Add Gaussian noise:

[0031] Here, the scalar hyperparameter belong Specifies the value range of the noise added in each iteration. As shown below:

[0032] Then, marginalize the intermediate steps: Here, To help the reverse diffusion process, additional information is used in the form of age features And optimize a neural denoising model that uses age features and the noisy target image As input,

[0033] Then, in the noise feature extraction stage, non-age features The convolutional layer is processed as follows:

[0034] Adding conditional constraints to the noise scheduling of the diffusion model significantly enhances the model's ability to control the synthetic content, ensuring that the generated samples are not only consistent with the statistical characteristics of the target distribution, but also highly correlated with specific conditions in various properties. As a constraint, the model can learn more information and thus generate diverse images.

[0035] In the denoising process, in addition to the source image , the denoising model will also is conditioned as input and trained to predict the noise vector .train The objective function of the process is:

[0036] The goal is to enhance the classifier's discriminative ability by spreading the generative process of the model. During the diffusion process, the classifier's decoupling ability is improved by continuously backpropagating the gradient, thereby improving the discriminative ability. Finally, the mean absolute error (MAE) loss is used:

[0037] in is the real age label. In addition, the deep neural network will also Output a 101-dimensional vector Therefore, the overall goal can be expressed as:

[0038] To further reduce the sorting and training time of the model, an active learning strategy is adopted, which aims to develop an efficient labeling algorithm by sampling the most representative queries. , this dataset is a face dataset from an unconstrained environment The data set is uniformly sampled and contains noise samples. Then, the sorted data set Used to train the dual-constraint diffusion model DCDM. The model is in the dataset After training, test the dataset without rearrangement and select samples containing noise labels. The selection strategy is as follows:

[0039] in is the label predicted by the deep neural network, Is a real age label. At the same time, Greater than The samples will be fed into the dataset , thus forming a new data set Then, the new dataset This is fed into a paired comparison algorithm. Through this process, the best model can be trained using as few samples as possible.

[0040] Feature distribution diagram A represents the feature distribution of the initial dataset samples, which contain noise samples and class imbalance. After processing by the model and algorithm, feature distribution diagram B is obtained. The noise samples are relabeled and the diffusion model generates a series of continuous face images to solve the class imbalance problem.

[0041] 2. Reasoning process: First, the trained and optimized model weights are loaded. Given an image x from the training set, the image features are extracted as F through the Resnet34 encoder based on the trained weights. The features are then decoupled into age features and non-age features. The age of the input image is predicted through the classifier output layer, thus achieving highly accurate and robust facial age estimation.

[0042] 3. Implementation Examples 1. Implementation Example 1: Construction and Training of Dual Constraint Diffusion Model DCDM (1) Composition: Data preprocessing: Align the input images using facial landmarks detection and resize them to a uniform size.

[0043] Feature Extractor: Uses the ResNet-34 network as the backbone network to extract features from the input image.

[0044] Diffusion model: Contains forward and backward processes to generate continuous age-progressive facial representations.

[0045] Conditional constraints: During the diffusion process, age-related and irrelevant features are used as conditional constraints to enhance the diversity and accuracy of generated samples.

[0046] (2) Effects: Accurate sample labeling: Through the paired comparison module, accurate labeling of samples is achieved, reducing the impact of noisy labels.

[0047] Continuous age sample generation: The diffusion model can generate a series of continuous, high-quality age samples, effectively solving the class imbalance problem.

[0048] Performance Improvement: Experimental results on multiple public datasets show that the dual-constrained diffusion model (DCDM) outperforms existing methods in facial age estimation tasks. 2. Implementation Example 2: Application of Active Learning Strategy (1) Composition: Model training: After preliminary training, the double-constrained diffusion model DCDM is used to test the unsorted dataset and select samples with larger prediction errors.

[0049] Sample selection strategy: Select samples that need to be relabeled based on the comparison results of the prediction error and the set threshold.

[0050] Dataset update: Add the selected re-labeled samples to the sorted dataset to form a new training dataset.

[0051] (2) Implementation process, such as Figure 2 shown 1) First, a small number of samples are selected from the unlabeled dataset and labeled to form a training set; 2) Select a training model and train it using a training set of a small number of labeled samples. This results in low performance. 3) Use the trained model to predict unlabeled samples; 4) Define a query strategy (including the uncertainty of the predictions and the query strategy applied to the requested annotations), return the priority scores of the unlabeled samples based on the strategy, and manually label the data that needs to be labeled; 5) Add the newly selected data to the training set to update the training set, and use the updated training set to train the model; 6) Determine whether the dual-constrained diffusion model (DCDM) meets the stopping criterion: If the stopping criterion is not met, continue to use the query strategy to select samples to be labeled and perform manual labeling, repeating steps 4-5-6 until the stopping criterion is met (this criterion can be the number of iterations or an indicator such as accuracy reaching a set value, such as model accuracy exceeding 99%), indicating that the model has been trained and can be used.

[0052] (3) Effects: Reduced labeling costs: Through active learning strategies, only samples with large prediction errors are labeled, reducing the required manual labeling work.

[0053] Improve model performance: By iteratively selecting and labeling the most representative samples, the classification effect of the model is improved.

[0054] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. An age estimation method based on a diffusion model, characterized in that: include: The dataset is obtained by uniformly sampling from a face dataset in an unconstrained environment containing noise samples; Inputting the data set into a facial age estimation model, wherein the facial age estimation model includes a paired comparison module and a double-constrained diffusion model connected in sequence, wherein the double-constrained diffusion model includes a backbone feature extractor and a diffusion model connected in sequence; A paired comparison module performs rearrangement labeling and noise removal to obtain an ordered, noise-free labeled face age image. A backbone feature extractor extracts features from the face age image and decouples the extracted features into additional variables including age features and non-age features. A forward diffusion process of a diffusion model gradually adds Gaussian noise to the ordered, noise-free labeled face age image over multiple iterations to obtain a noisy target image. The noisy target image and age features are subjected to neural denoising through the reverse generation process of the diffusion model to achieve aging reconstruction of the face age image and obtain the age corresponding to the face age image to obtain the trained facial age estimation model; The face image to be tested is input into the trained facial age estimation model to obtain the age corresponding to the face image to be tested.

2. The age estimation method based on the diffusion model according to claim 1, characterized in that: The paired comparison module is used to perform rearrangement labeling and noise removal to obtain an ordered face age image with a noise-free label, specifically including: Initialize the sample fraction of each sample in the dataset ,in, Indicates a score, represents a sample; Compare the samples in the dataset in pairs and score the image of the older individual in each pair; According to the cumulative scores of the samples, we continuously compare the samples with the same scores and iteratively improve the ranking of the samples to obtain the sorted data set. The formula is expressed as: ； in, Representative samples The score, Represents another sample The score, represents the sorted dataset, Represents a constant within (0, 1); The sorted dataset is first regularized, and the data samples are mapped to different age ranges according to the specific characteristics of each dataset; The mapped values ​​are rounded based on the following formula to obtain an ordered face age image with noise-free labels: ； in, represents the mean of the sorted data set, Represents the standard deviation of the sorted data set, interval Represents the lower and upper bounds of the sorted dataset.

3. The age estimation method based on the diffusion model according to claim 2, characterized in that: It also includes testing the unsorted dataset using a deep neural network after rearrangement labeling and noise removal through the paired comparison module, and selecting samples with large prediction errors based on the following formula: ； in, represents the label predicted by the deep neural network, Indicates the true age label; According to the comparison result between the prediction error and the set threshold, the re-labeled samples are selected, and the selected re-labeled samples are added to the sorted data set to obtain the data set.

4. The age estimation method based on the diffusion model according to claim 1, characterized in that: The forward diffusion process of the diffusion model gradually adds Gaussian noise to the ordered face age image with noise-free labels in multiple iterations to obtain a noisy target image, specifically including: In the forward process of the diffusion model, a forward Markov diffusion process guided by the non-age features in the additional variables is defined. The forward Markov diffusion process gradually adds Gaussian noise to the ordered face age images with noise-free labels in multiple iterations. The formula is expressed as: ； ； The intermediate steps are marginalized based on the following formula to obtain the noisy target image: ； in, represents the number of iterations, represents ordered and noise-free labeled face age images, represents a scalar hyperparameter, represents the sum of scalar hyperparameters over multiple iterations, Represents a face age image.

5. The age estimation method based on the diffusion model according to claim 1, characterized in that: The inverse generation process of the diffusion model is used to perform neural denoising on the noisy target image and age features to achieve aging reconstruction of the face age image and obtain the age corresponding to the face age image: In the reverse generation process of the diffusion model, the noisy target image and age features are used as the input of the neural denoising model, and the formula is expressed as: ； In the synthesis stage, the additional variables are guided as conditions, and the processing formula is expressed as: ； in, represents non-age characteristics in the additional variables, represents ordered and noise-free labeled face age images, represents the face age image, represents the noise vector.

6. The age estimation method based on the diffusion model according to claim 5, characterized in that: It also includes conditioning the age feature as the input of the neural denoising model to predict the noise vector; the objective function of the conditioning process is: ； During the diffusion process, the gradient is continuously back-propagated to improve the decoupling ability of the reverse process of the diffusion model, using the mean absolute error loss: ； in, Indicates the true age label; The overall objective function formula is expressed as: ； in, is the conditional process objective function, is the mean absolute error loss.

7. The age estimation method based on the diffusion model according to claim 1, characterized in that: The facial aging reconstruction specifically includes two facial aging reconstruction processes: a facial aging reconstruction process involving the same age but different identities and a facial aging reconstruction process involving the same identity but different ages; The reconstruction process is expressed as: ； ； in, and Represents two face images of the same age but different identities, and Two face images representing the same identity but different ages, and Respectively represent and age characteristics, and Respectively represent and Non-age characteristics, and Respectively represent and age characteristics, and Respectively represent and non-age characteristics.

Images