Age and gender prediction method based on multi-view image and application

Abstract

The invention relates to an age and gender prediction method based on a multi-view image and application. The method comprises the following steps: constructing an age and gender prediction model, and establishing an age prediction optimization model; a plurality of images of the face at different angles are collected and processed, and a training data set is established together with age and gender labels; inputting the training data into an age and gender prediction model for training, and optimizing an output result by an age prediction optimization model; after a plurality of images of different angles of the face are collected and processed, the images are input into the trained age and gender prediction model, the corresponding gender and a predicted age coarse value are output, the predicted age coarse value is optimized through an age prediction optimization model, and the optimized predicted age and gender are obtained; the method is used for predicting the age and the gender of a collected person after a face image is collected. According to the invention, the accuracy of age and gender prediction is significantly improved; dynamic adjustment is achieved, the accuracy of age prediction is continuously improved, and time and resource cost are greatly saved; and more personalized services can be conveniently provided for clients.

Description

Technical Field

[0001] This invention relates to the technical field of image or video recognition or understanding, and in particular to a method and application for age and gender prediction based on multi-view images. Background Technology

[0002] In today's digital age, age detection technology has wide applications in numerous fields, from security monitoring and human-computer interaction to beauty and healthcare. Accurate age detection provides crucial information for decision-making. Currently, age detection technology is primarily based on image analysis, and image analysis-based age detection techniques mainly focus on facial image analysis. One of the mainstream methods is facial feature extraction and age prediction using deep learning architectures such as Convolutional Neural Networks (CNNs). Some studies employ pre-trained CNN models, such as VGG16 or InceptionV3, and fine-tune them for age classification tasks, using facial features such as edges, shape, and texture to infer age. Under ideal conditions, such as uniform lighting, unobstructed faces, normal poses, and high dataset diversity, these methods exhibit a certain level of prediction accuracy. In some controlled experimental environments, the mean absolute error (MAE) of age prediction can be controlled within a certain range. However, existing models are highly dependent on the training dataset. If the dataset is biased, the model's generalization ability will be significantly reduced, and prediction accuracy will decrease markedly when facing different populations.

[0003] In the beauty and cosmetic industry, the shortcomings of existing age detection technologies are particularly prominent. Images collected by front-end beauty analysis instruments often differ significantly from the features of ideal training data due to factors such as makeup on the user's face, resulting in a severe decrease in the accuracy of image-based age detection.

[0004] Current age detection technology faces many challenges in practical applications, making it difficult to meet the needs of scenarios such as the beauty and cosmetic industry, which have high requirements for accuracy, real-time performance, and stability, and genuinely need to know the age of users. Innovative technologies are urgently needed to overcome these difficulties. Summary of the Invention

[0005] This invention solves the problems existing in the prior art and provides a method and application for age and gender prediction based on multi-view images.

[0006] The technical solution adopted in this invention is an age and gender prediction method based on multi-view images, wherein the method constructs an age and gender prediction model and establishes an age prediction optimization model;

[0007] Several images of the face from different angles are collected and processed, and a training data set is established together with age and gender labels; the training data is input into the age and gender prediction model for training, and the output results are optimized using the age prediction optimization model.

[0008] Several images of the face from different angles are collected and processed, then input into the trained age and gender prediction model. The model outputs the corresponding coarse values ​​of gender and predicted age. The coarse values ​​of predicted age are then optimized using the age prediction optimization model to obtain the optimized predicted age and gender.

[0009] Preferably, the age and gender prediction model includes a feature extraction module, a feature fusion module, and a fully connected layer arranged in sequence; the fully connected layer outputs a coarse predicted age and gender.

[0010] Preferably, the feature extraction module includes several convolutional branches arranged in parallel, and the number of convolutional branches is greater than or equal to the number of data points in each training data set.

[0011] Preferably, the age prediction optimization model includes an input layer, a similarity comparison module, a regression calculation module, a judgment layer, and an output layer arranged in sequence, wherein the judgment layer conditions are returned to the input end of the similarity comparison module; and a storage unit is provided in conjunction with the regression calculation module.

[0012] Preferably, the fused features output by the feature fusion module and the corresponding predicted age are input into the input layer; the similarity comparison module is equipped with a standard feature library and outputs two standard feature vectors corresponding to the two sides of the predicted age with the highest similarity.

[0013] Preferably, the fused features output by the feature fusion module and two standard feature vectors are input into the regression calculation module, which outputs the optimized undetermined age and submits it to the judgment layer for judgment.

[0014] Preferably, if the judgment layer determines the output, it obtains the average value of the iterative difference recorded in the storage unit and selects the output path of the output layer based on a preset strategy.

[0015] Preferably, any set of images acquired includes a frontal image and two side images.

[0016] Preferably, the two side images are acquired synchronously, and the angle between the acquisition angle and the acquisition direction of the front image is less than 90°.

[0017] An application of age and gender prediction based on multi-view images is used to predict the age and gender of the person being photographed after facial images are captured.

[0018] This invention relates to a method and application for age and gender prediction based on multi-view images. The method constructs an age and gender prediction model and establishes an age prediction optimization model; it acquires and processes several images of the face from different angles, and establishes a training data set together with age and gender labels; it inputs the training data into the age and gender prediction model for training, and optimizes the output results using the age prediction optimization model; it acquires and processes several images of the face from different angles, inputs them into the trained age and gender prediction model, outputs the corresponding gender and coarse predicted age values, and optimizes the coarse predicted age values ​​using the age prediction optimization model to obtain the optimized predicted age and gender; the method is applied to predict the age and gender of the person being captured after acquiring facial images.

[0019] The beneficial effects of this invention are as follows:

[0020] (1) By integrating multi-view information from 2D and 3D UV images, the complementary features in different modal data can be fully explored, significantly improving the accuracy of age and gender prediction, and providing more accurate customer profiles for the beauty industry and other fields.

[0021] (2) The unique adaptive age prediction optimization mechanism enables the model to dynamically adjust according to the prediction error, automatically include the accurately predicted data into the adaptive data pool, and continuously improve the accuracy of age prediction by assisting prediction with similar data, without the need to retrain the model, thus saving time and resource costs significantly.

[0022] (3) It is particularly suitable for the beauty industry. As a reliable and efficient customer information analysis tool, it can provide customers with more personalized services by accurately predicting age and gender, including customized solutions and accurate product recommendations, which helps to improve service quality and market competitiveness. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention;

[0024] Figure 2 This is a schematic diagram of the age and gender prediction model of the present invention;

[0025] Figure 3 This is a schematic diagram illustrating the application of the age and gender prediction model in this invention.

[0026] Figure 4 This is a schematic diagram of the age prediction optimization model of the present invention;

[0027] Figure 5 This is a schematic diagram illustrating the overall implementation of the method of the present invention. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to embodiments, but the scope of protection of the present invention is not limited thereto.

[0029] This invention relates to a method for predicting age and gender based on multi-view images. It is generally applied in a medium or device. The medium stores the multi-view image-based age and gender prediction program, which, when executed by a processor, implements the method. The device is typically a front-end device used in the beauty and cosmetic industry, comprising an image acquisition unit, a memory, a processor, and a computer program stored in the memory and executable on the processor. Images acquired by the image acquisition unit are transmitted to the processor, and when the processor executes the program, it implements the aforementioned multi-view image-based age and gender prediction method.

[0030] The concept of this invention is to construct an adaptive data processing mechanism that continuously optimizes the accuracy of age prediction without retraining the model. By deeply mining and analyzing multi-view data, the features of the collected images are dynamically integrated and analyzed. Combined with real-time feedback and historical data, the prediction strategy is continuously adjusted, thereby achieving a continuous improvement in the accuracy of age prediction.

[0031] The method of the present invention includes the following steps:

[0032] (1) Construct age and gender prediction models and establish age prediction optimization models;

[0033] (2) Collect and process several images of the face from different angles, and establish a training data set together with age and gender labels; input the training data into the age and gender prediction model for training, and optimize the output results using the age prediction optimization model.

[0034] (3) After collecting and processing several images of the face from different angles, input them into the trained age and gender prediction model, output the corresponding gender and predicted age coarse value, and optimize the predicted age coarse value with the age prediction optimization model to obtain the optimized predicted age and gender.

[0035] First, both training and practical applications require the acquisition and processing of several images of the face from different angles. Each set of acquired images includes a frontal image and two side images.

[0036] The two side images are acquired synchronously, and the angle between the acquisition angle and the acquisition direction of the front image is less than 90°.

[0037] Taking beauty instruments as an example, they simultaneously capture three images of the face: a frontal image and two side images, left and right. The acquisition angle of the side images is generally 30° to 60° from the frontal image, and this angle can be adjusted based on the facial features to be analyzed.

[0038] After acquiring three images, the processing involves using facial feature extraction algorithms to accurately identify the edges between the face (foreground) and the background, segmenting the facial region in the image, and then using a normalization algorithm to adjust the pixel values ​​of each group of three images. Mapped to a specific interval, satisfying

[0039]

[0040] in, These are the pixel values ​​of the original image. and These are the minimum and maximum pixel values ​​of the image, respectively, which enhances the consistency and comparability of image features.

[0041] The processing also includes resizing each group of three images after normalization, converting them to a preset dimension, such as 320*320, so that the input dimension data of the images matches the input of the model.

[0042] For the acquired images, this invention constructs age and gender prediction models and establishes an optimized age prediction model.

[0043] The age and gender prediction model includes a feature extraction module, a feature fusion module, and a fully connected layer arranged in sequence; the fully connected layer outputs a coarse predicted age and gender.

[0044] The following description is based on specific examples.

[0045] The feature extraction module includes several convolutional branches arranged in parallel, and the number of convolutional branches is greater than or equal to the number of data points in each training data set.

[0046] To deeply mine the complementary information in each group of three images, several convolutional branches set up in parallel are used to extract features from the preprocessed images of the face from the left, center, and right angles respectively:

[0047] Each convolutional branch includes multiple convolutional layers; for the left image, this is achieved through multiple convolutional layers. , ,..., Feature extraction is performed using the following formula: (k = 1,2,...,n, (The left image is the input image), to obtain the features of the left image. Similarly, the frontal image and the other side image are processed to obtain the features of the frontal image. and right-side image features To ensure alignment of the three features, each convolutional branch has the same number of convolutional layers. The resulting three features are then fused; this fusion refers to directly concatenating these three features along the channel dimension to obtain the fused feature. ;

[0048] The advantage of this method of extracting and fusing features from different angles is that:

[0049] (1) It can effectively preserve the specificity of features and fully extract the unique features contained in different side images of the face that are valuable for age and gender prediction. For example, the left image may highlight the contour lines and specific texture information on the left side of the face, the middle image focuses on the frontal details of the facial features, and the right image shows the features of the right side of the face. Compared with the method of first stitching and then extracting features, this operation can avoid the mutual interference of features from different sides during the convolution process. Moreover, each convolution branch can learn the features of different angles of the image based on the needs, thereby capturing these unique features more accurately.

[0050] (2) Eliminate adaptive differences and avoid deviations in feature recognition results due to factors such as acquisition angle and / or lighting conditions. By adaptively adjusting the parameters of each convolution channel according to the specific situation of the corresponding image, the feature extraction effect can be better adapted to the characteristics of different images.

[0051] (3) The model is highly flexible and scalable. In the subsequent model improvement process, if it is necessary to optimize the feature extraction of a certain side image, such as adjusting the depth of the convolutional layer or changing the convolutional kernel, only the corresponding convolutional channel needs to be modified, which will not affect the feature extraction of other side images.

[0052] The age and gender prediction model obtains fused features through a feature extraction module and a feature fusion module. After passing through a fully connected layer, the model outputs a coarse predicted age and gender based on different activation functions. Generally, a linear activation function is used for age prediction, while a sigmoid activation function is used for gender classification. During training, the accurately labeled age and gender information are used as labels, and a loss function combining weighted cross-entropy loss and mean squared error loss is used to optimize the age and gender prediction model. The loss function satisfies the following conditions:

[0053]

[0054] in, Cross-entropy loss for gender prediction, This represents the loss due to the mean squared error in age prediction. , The weights are used to continuously adjust the model parameters through the backpropagation algorithm to minimize the loss value of the model on the training set, thereby obtaining a high-performance age and gender prediction model.

[0055] It should be noted that the gender prediction is a binary classification prediction, which must be accurate. However, the output coarse predicted age value needs to be aligned and optimized for age prediction.

[0056] To address this, the present invention proposes an optimized age prediction model, comprising an input layer, a similarity comparison module, a regression calculation module, a judgment layer, and an output layer arranged sequentially. The judgment layer returns the conditions to the input of the similarity comparison module; a storage unit is provided in conjunction with the regression calculation module.

[0057] Further, the following description is provided in conjunction with the embodiments.

[0058] The fused features output by the feature fusion module and the corresponding predicted age are input into the input layer; the similarity comparison module has a standard feature library and outputs two standard feature vectors corresponding to the two sides with the maximum similarity between the predicted age.

[0059] In this invention, the standard feature library is established based on historical data. After obtaining the feature vector of each data point by inputting an image, an initial feature library is constructed. The extracted features are associated with and stored with the corresponding age information to form a structured data set, providing data support for subsequent prediction optimization. At the same time, the feature library is updated with new features and corresponding ages to serve as the standard feature library. The update frequency can be set manually, such as real-time updates, direct updates based on effective detection data, or updates on a daily or weekly basis.

[0060] With fusion features The final feature vector of the image in the first stage is compared with the predicted age, such as 29 years old, and then compared with data in the standard feature library. The similarity comparison module calculates the input features. The data is selected based on the cosine similarity between the data and the features in the standard feature library. The data is selected based on the highest feature similarity when the age is less than or equal to (e.g., 14, 20, 25 years old) and greater than (e.g., 29, 35, 42 years old) the predicted age input.

[0061] The fused features and two standard feature vectors output by the feature fusion module are input into the regression calculation module, which outputs the optimized undetermined age and submits it to the judgment layer for judgment.

[0062] Through the above operations, a total of two relevant feature layer data and their corresponding age information were obtained. Based on these three data (input image features) Characteristics of younger age and characteristics of older age Regression calculations are performed to obtain new predicted age values;

[0063] Specifically, regression calculations can employ methods such as weighted averaging, assigning weights based on feature similarity, using the following formula:

[0064]

[0065] in, Here are the weighting coefficients, i = 1, 2, 3 and =1, the weight is determined based on feature similarity. Similarity to other feature vectors (including (itself), and ensures that the total value of w is 1 through softmax;

[0066] This represents the similarity weight between the predicted age and the predicted age. The similarity weights between predicted age and younger age are represented. The similarity weights represent the predicted age and the actual age; a softmax() operation is performed on these three factors to make the sum of their similarities equal to 1, representing the weights of the distance between the current feature and the features of youth, itself, and age, which are used to calculate the age in the next round.

[0067] Calculated , , The new age is obtained by weighting the corresponding age value. ,

[0068]

[0069] Update the current input data features as follows To adapt to new age ,

[0070]

[0071] The updated data features are then passed to the decision layer for evaluation. If the output requirements are not met, the updated data features replace the features from the previous iteration and are input into the input layer. This process is repeated until the difference between the newly predicted age and the previously predicted age is less than a preset threshold T (generally, T is 2), or the number of iterations reaches a set upper limit. At this point, the final age is used as the predicted age of the input image.

[0072] If the judgment layer determines the output, it obtains the average value of the iterative difference recorded in the storage unit and selects the output path of the output layer based on the preset strategy.

[0073] Specifically, the decision layer determines the data processing method by monitoring the relationship between the interpolated average of the new age prediction in each iteration and the threshold, thereby ensuring the accuracy and stability of the model. The specific steps are as follows:

[0074] For the i-th iteration, let the new age prediction value be... The age prediction value from the previous iteration was The difference in this iteration is As the number of iterations increases, the average value of these interpolations is calculated. where n is the current iteration number;

[0075] A reasonable threshold T is preset; when the calculated interpolated average value is... If the prediction is positive, it means that the fluctuation of the age prediction value is small during the iteration process, the prediction result of the model is relatively stable, and the reliability of the new prediction data is high. Therefore, the features of the input image and the corresponding age and gender information are directly added to the standard feature library to provide more references for the model's subsequent predictions and help to further optimize the model's performance. Otherwise, it indicates that the age prediction value changes significantly during the iteration process, the stability of the model is affected, and the reliability of the new prediction data is questionable. In this case, proceed to the next step.

[0076] The data is marked as pending review and submitted for manual review. The manual review includes assessing image quality, such as whether there are problems such as blurring, occlusion, and uneven lighting, to avoid affecting the model's feature extraction and prediction accuracy. It also includes judging the reasonableness of the predicted age with the actual situation. Only after the accuracy and reliability of the data are confirmed by manual review will a decision be made on whether to add it to the standard feature library, so as to ensure that the data added to the feature library can play a positive role in improving the model's performance.

[0077] The present invention also relates to an application of the aforementioned age and gender prediction method based on multi-view images, which is used to predict the age and gender of the person being photographed after facial images are acquired.

[0078] In this invention, the method is implemented based on a device, such as a facial image acquisition device. By acquiring images from different angles (not necessarily frontal and side images), an initial set of three images can be obtained. The acquired images are fed back to a computer device, where the controller outputs the coarse values ​​of gender and predicted age corresponding to the acquired images using a trained age and gender prediction model. The coarse value of predicted age is then optimized using an age prediction optimization model to obtain the optimized predicted age and gender before being output.

[0079] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0080] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0082] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0083] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0084] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting age and gender based on multi-view images, characterized in that: The method constructs age and gender prediction models and establishes an age prediction optimization model. Collect and process several images of the face from different angles, and build a training dataset together with age and gender labels; The age and gender prediction model is trained by inputting training data, and the output is optimized using the age prediction optimization model. Several images of the face from different angles are collected and processed, then input into the trained age and gender prediction model. The model outputs the corresponding coarse values ​​of gender and predicted age. The coarse values ​​of predicted age are then optimized using the age prediction optimization model to obtain the optimized predicted age and gender.

2. The age and gender prediction method based on multi-view images according to claim 1, characterized in that: The age and gender prediction model includes a feature extraction module, a feature fusion module, and a fully connected layer arranged in sequence; the fully connected layer outputs a coarse predicted age and gender.

3. The age and gender prediction method based on multi-view images according to claim 2, characterized in that: The feature extraction module includes several convolutional branches arranged in parallel, and the number of convolutional branches is greater than or equal to the number of data points in each training data set.

4. The age and gender prediction method based on multi-view images according to claim 2, characterized in that: The age prediction optimization model includes an input layer, a similarity comparison module, a regression calculation module, a judgment layer, and an output layer arranged in sequence. The judgment layer returns the conditions to the input of the similarity comparison module. A storage unit is provided in conjunction with the regression calculation module.

5. The age and gender prediction method based on multi-view images according to claim 4, characterized in that: The fused features output by the feature fusion module and the corresponding predicted age are input into the input layer; The similarity comparison module is equipped with a standard feature library and outputs two standard feature vectors corresponding to the two sides with the highest similarity result of the predicted age.

6. The age and gender prediction method based on multi-view images according to claim 5, characterized in that: The fused features and two standard feature vectors output by the feature fusion module are input into the regression calculation module, which outputs the optimized undetermined age and submits it to the judgment layer for judgment.

7. The age and gender prediction method based on multi-view images according to claim 4, characterized in that: If the judgment layer determines the output, it obtains the average value of the iterative difference recorded in the storage unit and selects the output path of the output layer based on the preset strategy.

8. The age and gender prediction method based on multi-view images according to claim 1, characterized in that: Each set of images acquired includes a frontal image and two side images.

9. The age and gender prediction method based on multi-view images according to claim 8, characterized in that: The two side images are acquired synchronously, and the angle between the acquisition angle and the acquisition direction of the front image is less than 90°.

10. An application of the age and gender prediction method based on multi-view images as described in any one of claims 1 to 9, characterized in that: It is used to predict the age and gender of the person after facial images are captured.

Images