Face image generation method and device, computer equipment and storage medium

Abstract

The invention relates to a face image generation method and apparatus, a computer device and a storage medium. The method comprises the steps of obtaining a target genetic image of an object having a genetic relationship with a target object; determining a target genetic feature corresponding to each structural unit according to the target genetic image; using a pre-trained feature analysis model to analyze the target genetic feature corresponding to each structural unit, and determining the target genetic feature; and calling a pre-trained face image generation model, processing the target genetic features, and generating a face image of the target object.

Description

Technical Field

[0001] This application relates to the field of genetic analysis technology, specifically to a method, apparatus, computer device, and storage medium for generating human face images. Background Technology

[0002] In today's society, the rapid development of technology has brought tremendous changes to various fields. Among them, image recognition and processing technology plays a crucial role in many scenarios, and has had a profound impact on the acquisition and utilization of facial images in many aspects.

[0003] However, handling cases involving the malicious transfer of underage children presents unique challenges. In these cases, accurate photographs of the children play a crucial role in quickly locating them and confirming their identities. However, in reality, parents often fail to provide clear and accurate photographs of their children for various reasons. Some families may have very few photos of their children due to limited living conditions. Therefore, obtaining images that truly reflect the child's appearance under such circumstances becomes a critical issue that urgently needs to be addressed in handling these cases. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for generating facial images. The method can generate a facial image of a target object based on a target kinship image. The generated facial image can highly realistically reflect the current facial features of a maliciously transferred young person, providing crucial and powerful evidence for handling such cases.

[0005] A first aspect of this application provides a method for generating a face image, the method comprising: Obtain the target kinship image of objects that are related to the target object by blood; Based on the target kinship image, determine the target kinship features corresponding to each structural unit; Using a pre-trained feature analysis model, the target kinship features corresponding to each structural unit are analyzed to determine the target genetic features; A pre-trained face image generation model is invoked to process the target genetic features and generate a face image of the target object.

[0006] Optionally, the step of using a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features includes: The pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, and the first feature corresponding to each structural unit is obtained. For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the first feature corresponding to the structural unit; The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0007] Optionally, the step of using a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features includes: For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the target kinship feature corresponding to the structural unit; A pre-trained feature analysis model is invoked to process the discrete features corresponding to each structural unit, thereby determining the target genetic feature.

[0008] Optionally, the step of using a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features includes: A pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, thereby obtaining a probability set corresponding to each structural unit, wherein the probability set includes the probability corresponding to each category; For each structural unit, determine the category corresponding to the highest probability in the corresponding probability set, and determine the discrete features corresponding to the structural unit by using the discrete features corresponding to the category. The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0009] Optionally, the pre-trained feature analysis model includes a sub-attention module and a feature processing module. The step of using the pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features includes: Using the sub-attention module, the target affinity features corresponding to each structural unit are analyzed to obtain the second feature; The second feature is processed using the feature processing module to obtain the target genetic feature.

[0010] Optionally, the step of calling a pre-trained face image generation model to process the target genetic features and generate a face image of the target object includes: Obtain the attribute information of the target object; A pre-trained face image generation model is invoked to process the target genetic features and attribute information to generate a face image of the target object.

[0011] Optionally, determining the target kinship features corresponding to each structural unit based on the target kinship image includes: The target kinship image is divided to obtain a unit image corresponding to each structural unit; For each structural unit, the feature extraction model corresponding to the structural unit is obtained. The feature extraction model is used to process the unit image corresponding to the structural unit to obtain the target kinship feature corresponding to the structural unit.

[0012] Optionally, the method further includes: Obtain the target description text; When the target description feature is not extracted from the target description text, the feature analysis model is invoked to process the target kinship feature to obtain the target genetic feature of the target object; the target description feature is the facial description feature of the target object; When the target description feature is extracted from the target description text, it is detected whether the target description feature includes genetic description feature and face description feature, and the feature in the detection result is obtained. The genetic feature analysis module is invoked to process the target kinship features and the features in the detection results to obtain the face image of the target object.

[0013] In a second aspect, this application provides a face image generation apparatus, comprising: The acquisition unit is used to acquire target kinship images of objects that are related to the target object by blood. The first determining unit is used to determine the target kinship features corresponding to each structural unit based on the target kinship image; The second determining unit is used to analyze the target kinship features corresponding to each structural unit using a pre-trained feature analysis model to determine the target genetic features; The calling unit is used to call a pre-trained face image generation model to process the target genetic features and generate a face image of the target object.

[0014] Optionally, the second determining unit is used for: The pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, and the first feature corresponding to each structural unit is obtained. For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the first feature corresponding to the structural unit; The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0015] Optionally, the second determining unit is used for: For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the target kinship feature corresponding to the structural unit; A pre-trained feature analysis model is invoked to process the discrete features corresponding to each structural unit, thereby determining the target genetic feature.

[0016] Optionally, the second determining unit is used for: A pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, thereby obtaining a probability set corresponding to each structural unit, wherein the probability set includes the probability corresponding to each category; For each structural unit, determine the category corresponding to the highest probability in the corresponding probability set, and determine the discrete features corresponding to the structural unit by using the discrete features corresponding to the category. The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0017] Optionally, the pre-trained feature analysis model includes a sub-attention module and a feature processing module, and the second determining unit is used for: Using the sub-attention module, the target affinity features corresponding to each structural unit are analyzed to obtain the second feature; The second feature is processed using the feature processing module to obtain the target genetic feature.

[0018] Optionally, the calling unit is used for: Obtain the attribute information of the target object; A pre-trained face image generation model is invoked to process the target genetic features and attribute information to generate a face image of the target object.

[0019] Optionally, the first determining unit is used for: The target kinship image is divided to obtain a unit image corresponding to each structural unit; For each structural unit, the feature extraction model corresponding to the structural unit is obtained. The feature extraction model is used to process the unit image corresponding to the structural unit to obtain the target kinship feature corresponding to the structural unit.

[0020] A third aspect of this application provides a computer device, including: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0021] A fourth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the method as described in any of the above.

[0022] In this embodiment, a target kinship image of a blood relative to the target object is obtained; based on the target kinship image, the target kinship features corresponding to each structural unit are determined; a pre-trained feature analysis model is used to analyze the target kinship features corresponding to each structural unit to determine the target genetic features; a pre-trained face image generation model is invoked to process the target genetic features and generate a face image of the target object. This application can generate a face image of the target object based on the target kinship image. The generated face image can highly realistically reflect the current facial features of the maliciously transferred young object, providing crucial and powerful evidence for handling such cases. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a face image generation method provided in one embodiment of this application; Figure 2 A flowchart illustrating a method for determining target genetic characteristics provided in one embodiment of this application; Figure 3 A flowchart illustrating a method for determining target genetic characteristics provided in one embodiment of this application; Figure 4 A flowchart illustrating a method for determining target genetic characteristics provided in one embodiment of this application; Figure 5 A bar chart showing the number of cases involving the malicious transfer of young victims solved in the middle of the period for the method of determining the target genetic characteristics of this application; Figure 6 The number of times the ranking of a 2 million-repository library improved under different original library ranking intervals in the method for determining the target genetic characteristics of this application; Figure 7 This is a schematic diagram of a face image generation device provided in one embodiment of this application; Figure 8 This is a schematic diagram of a computer device structure provided in one embodiment of this application. Detailed Implementation

[0024] In today's society, the rapid development of technology has brought about tremendous changes in various fields. Among them, image recognition and processing technology plays a crucial role in many scenarios, and its acquisition and utilization of facial images has had a profound impact on many aspects. However, when dealing with cases such as the malicious transfer of young children, unique challenges arise. In these cases, accurate photos of the children are crucial for quickly locating them and confirming their identities. But in reality, parents may be unable to provide clear and accurate photos of their children for various reasons. Some families may have very few photos of their children due to limited living conditions. Therefore, how to obtain images of children that truly reflect their appearance under the predicament of lacking accurate photos has become a critical issue that urgently needs to be addressed in the handling of such cases. To address the aforementioned problems, this application provides a method for generating facial images. The method involves acquiring a target kinship image of a blood relative of the target object; determining the target kinship features corresponding to each structural unit based on the target kinship image; using a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit to determine the target genetic features; and calling a pre-trained facial image generation model to process the target genetic features to generate a facial image of the target object. This application can generate a facial image of the target object based on the target kinship image. The generated facial image can highly realistically reflect the current facial features of a maliciously transferred young person, providing crucial and powerful evidence for handling such cases.

[0025] The solutions in this application embodiment can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0026] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0027] Please see Figure 1 The following embodiments use a face generation system as the execution subject, applying the method provided in the embodiments of this application to the aforementioned face generation system. The face image generation method provided in the embodiments of this application includes the following steps 101-104: Step 101: Obtain the target kinship image of the object that is related to the target object by blood.

[0028] Among them, the target kinship images are facial images of people who are related to the target subject by blood, which can cover the target subject's immediate family members, such as parents, children, and siblings. These images can be obtained through various means, such as using professional image acquisition equipment to photograph people who are related to the target subject; or by selecting suitable images from an existing family photo library.

[0029] Meanwhile, to ensure the effectiveness of subsequent analysis, the collected images of the target relatives should have a certain quantity and diversity. Sufficient quantity can provide rich data support for the analysis, while diversity is reflected in the fact that the images include different age groups, different expressions, and different shooting angles, which helps to comprehensively capture the expression of genetic characteristics under different conditions.

[0030] Step 102: Determine the target kinship features corresponding to each structural unit based on the target kinship image.

[0031] Among them, the target kinship features include facial feature information of objects that are related to the target object by blood, such as facial contour features, facial features, etc., which describe the characteristics of the object's face from various aspects.

[0032] In this step, the human face can typically be divided into multiple structural units, such as eyes, nose, mouth, and cheeks, each carrying unique genetic information. Therefore, based on the principles of facial anatomy and visual structure division, the entire face can be divided into sub-images corresponding to multiple structural units. Examples include hair, forehead, left eye, right eye, left ear, right ear, nose, left cheek, right cheek, mouth, and chin. Then, for each structural unit, its corresponding feature extraction model and sub-image are obtained. The feature extraction unit processes the sub-image corresponding to that structural unit to obtain the target kinship features. For example, for the eye structural unit, its corresponding feature extraction unit can use edge detection algorithms to determine the eye contour, and then extract geometric features such as palpebral fissure length, interocular distance, and eyelid folds; simultaneously, it uses color analysis algorithms to obtain color and morphological features such as iris color and pupil size. For the nose structural unit, its corresponding feature extraction unit uses 3D reconstruction technology to obtain three-dimensional features such as nasal bridge height, nasal tip angle, and nasal wing width; and uses texture analysis algorithms to extract the texture details of the nasal surface.

[0033] Alternatively, a first deep neural network is used to process the face image to obtain a feature map corresponding to the face image. Simultaneously, the face image is segmented to obtain a set of face region masks, which includes the face region mask corresponding to each structural unit. Then, the face region masks in the face region mask set are scaled to the same size as the aforementioned feature map, and the scaled masks are binarized to obtain binarized features, which in turn create a binarized feature set. For each binary feature corresponding to a structural unit in the binarized feature set, this binarized feature is multiplied by the aforementioned feature map to obtain a multiplication result. This multiplication result is then input into a second deep neural network to obtain the feature corresponding to that structural unit.

[0034] It should be noted that the neural network described above is trained on a large amount of data. It can be trained together with the models discussed later, or it can be trained using other methods. This is not a limitation here.

[0035] Based on the above, when there is only one target kinship image, the target kinship features corresponding to each structural unit can be obtained. When there are multiple target kinship images, one object may correspond to multiple images. Therefore, these target kinship images are divided into multiple image sets, each containing images of the same object. Then, for each image set, the features corresponding to each image in the set are obtained using the feature extraction method described above. These features are then comprehensively analyzed to obtain the final features corresponding to the set. Finally, each final feature is combined with the corresponding final features to obtain the target kinship features corresponding to each structural unit.

[0036] Step 103: Using a pre-trained feature analysis model, analyze the target kinship features corresponding to each structural unit to determine the target genetic features.

[0037] The target genetic features include facial features of the target subject, such as facial contours and facial features, which characterize the subject's face from various aspects. The pre-trained feature analysis model is built based on a large amount of facial genetic data and machine learning algorithms. This model aims to extract deep genetic information from the target kinship features corresponding to each structural unit, thereby determining the target genetic features.

[0038] In this step, when analyzing the target kinship characteristics, the model performs a comprehensive analysis of the target kinship characteristics of each structural unit. Through this comprehensive and in-depth analysis, the model ultimately determines the target genetic characteristics that can represent the genetic traits of the target object.

[0039] To extract target genetic features more accurately, this application can set different feature analysis models for different combinations of kinship relationships. This involves first determining the kinship relationship between each object and the target object, and then determining the feature analysis model to be used based on the combination of these kinship relationships. For example, one feature analysis model corresponds to the father and mother, and another to the father and his siblings.

[0040] For example, when the father and mother each correspond to a feature analysis model, the method for obtaining the pre-trained feature analysis model is as follows: Assume the father's face region encoding feature is F_r, the mother's face region encoding feature is M_r, and the offspring's face region encoding feature to be predicted is S_r. Select several parent-child pairing data to construct a training dataset with the parents' encoding features as input and the offspring's encoding features as labels. Concatenate the father's face region encoding feature F_r and the mother's face region encoding feature M_r to obtain a fused feature P_r. Input the fused feature into the feature analysis model, and the output is the probability vector V_r for each region of the offspring. Perform One-Hot encoding on the offspring's face region encoding feature S_r to obtain labels. Train the feature analysis model by calculating the cross-entropy loss between the probability vector V_r and the One-Hot encoded labels. Finally, the pre-trained feature analysis model is obtained.

[0041] In the above process, if the target description text can still be obtained, which may include facial description information of the target object, then it is detected whether the target description feature is extracted from the target description text. This target description feature is the facial description feature of the target object. If the target description feature is not extracted from the target description text, the feature analysis model is called to process the target kinship features to obtain the target genetic features of the target object. If the target description feature is extracted from the target description text, it is detected whether the target description feature includes both genetic description features and facial description features. The features in the detection results are obtained, and the genetic feature analysis module is called to process the target kinship features and the features in the detection results to obtain the facial image of the target object.

[0042] Among them, genetic descriptive features are heritable facial characteristics within the target individual's family, such as common family features like a high nose bridge, deep-set eyes, and specific ear shapes. It also includes descriptions of physiological characteristics with heritable predispositions, such as skin color, hair color, and eye color. Facial descriptive features are unique facial details of the target individual, such as the location and shape of moles, birthmarks, and dimples, as well as facial expression habits, such as the upward curve of the corners of the mouth when smiling and the lines formed when frowning.

[0043] When the detection result indicates that the target descriptive features include both genetic descriptive features and facial descriptive features, the feature analysis module can further include a generation module and a modification model. The specific steps of calling the feature analysis module to process the target kinship features and the features in the detection result to obtain the facial image of the target object are as follows: The generation module can obtain the weights corresponding to the target kinship features and the genetic descriptive features, and based on these two weights, fuse the target kinship features and the genetic descriptive features to obtain a first fused feature. The first fused feature is then processed to obtain a reference genetic feature. The image modification model obtains the weights corresponding to the facial descriptive features and the reference genetic feature, and based on these two weights, fuses the facial descriptive features and the reference genetic feature to obtain a second fused feature. The second fused feature is then processed to obtain the target genetic feature.

[0044] The fusion process described above can involve multiplying the features and their corresponding weights, and then adding the results to obtain the fused features. Other fusion methods can also be used, and this is not limited here.

[0045] In the above process, when there are multiple target kinship features, the kinship relationship between the target kinship feature and the target object can be obtained, the weight corresponding to each kinship object can be determined, and the target kinship feature and the corresponding weight can be input into the feature analysis model so that the genetic feature analysis model can analyze and process the input data to obtain the target genetic feature.

[0046] The aforementioned feature analysis model is built upon a large amount of facial genetic data and machine learning algorithms. This model aims to extract deep-seated genetic information from the target kinship features corresponding to each structural unit, thereby determining the target genetic characteristics.

[0047] When the kinship relationship between the kinship object and the target object is the first kinship relationship, the weight corresponding to the kinship object is set as the first weight; when the kinship relationship between the kinship object and the target object is the second kinship relationship, the weight corresponding to the kinship object is set as the second weight; when the kinship relationship between the kinship object and the target object is the third kinship relationship, the remaining total weight is determined according to the first and second weights; the ratio of the remaining total weight to the number is determined as the third weight, and the weight corresponding to the kinship object is set as the third weight.

[0048] The first kinship relationship is the relationship between the mother and the target. The second kinship relationship is the relationship between the father and the target. The third kinship relationship is the relationship between the siblings and the target.

[0049] In this step, for each related object, when the kinship between the related object and the target object is the first kinship, the weight corresponding to the related object is set as the first weight. When the kinship between the related object and the target object is the second kinship, the weight corresponding to the related object is set as the second weight. When the kinship between the related object and the target object is the third kinship, the remaining total weight is determined based on the first and second weights. Subsequently, the ratio of the remaining total weight to the corresponding number is calculated to determine the third weight, and the weight corresponding to the related object is set as the third weight. The specific steps for determining the remaining total weight based on the first and second weights are as follows: since the sum of all weights is 1, 1 is subtracted from the first weight and the second weight, and the difference is determined as the remaining total weight.

[0050] For example, if the weight corresponding to the first kinship is 0.3, the weight corresponding to the second kinship is 0.3, the weight corresponding to the third kinship is 0.35, and the number of kinship objects corresponding to the third kinship is 5, then the ratio of the third weight of 0.35 to 5 is 0.7.

[0051] Alternatively, statistical methods can be used to process target kinship features and determine the target genetic features of the target object. The specific steps are as follows: For each structural unit, in the genetic feature database, obtain the conditional probabilities of all first genetic feature combinations under the target kinship feature corresponding to that structural unit, obtain the conditional probability corresponding to each first genetic feature combination, and then determine the first genetic feature combination with the highest conditional probability and identify it as the second genetic feature combination; based on the conditional probability corresponding to the second genetic feature combination and the target kinship feature, determine whether the feature in the second genetic feature combination is a strong genetic feature; when the feature in the second genetic feature combination is a strong genetic feature, identify the feature in the second genetic feature combination as the target genetic feature; when the feature in the second genetic feature combination is not a strong genetic feature, find the genetic feature combination with the highest correlation to the second genetic feature combination based on the genetic feature matrix, and identify the feature in the genetic feature combination as the target genetic feature.

[0052] The genetic feature matrix includes the correlation between combinations of genetic features, which are determined using genetic statistics methods. The genetic feature database stores all features, all combinations of genetic features, and the conditional probability of each combination of genetic features for each feature. A combination of genetic features is a combination of multiple genetic features.

[0053] The specific steps for determining whether a feature in the second genetic feature combination is a strong genetic feature combination based on the conditional probability corresponding to the second genetic feature combination and the target kinship feature are as follows: detect whether the target kinship feature is in the preset strong feature set and whether the conditional probability corresponding to the second genetic feature combination is greater than the dominant inheritance threshold. When both conditions are met, the feature in the second genetic feature combination is determined to be a strong genetic feature; otherwise, the feature in the second genetic feature combination is not a strong genetic feature.

[0054] For example, target kinship characteristics

[0055] in, Indicates the first The kinship feature corresponding to each structural unit (hereinafter referred to as the i-th kinship feature), where n is the total number of structural units.

[0056] For the i-th kinship trait, its corresponding combination of second genetic traits The corresponding feature combinations, where It is a set containing all combinations of features. For the j-th combination of genetic traits, Obtain the conditional probability corresponding to the j-th genetic feature combination under the i-th kinship feature.

[0057] Next, it checks whether the i-th kinship feature belongs to the strong feature set D, and then checks the second genetic feature combination. If the corresponding conditional probability is greater than the dominant inheritance threshold, then the second genetic trait is combined. The characteristics identified in the text are strong heritable traits. Otherwise, the genetic trait combination with the strongest association with the second genetic trait combination will be selected, and the genetic trait in that combination will be identified as the target genetic trait. Specifically, the formula is as follows:

[0058] in, The combination of genetic characteristics in the genetic transition matrix Combination of genetic traits The strength of the association between two genetic traits is the probability / correlation of a combination of genetic traits transforming into, retaining, or co-occurring with another combination of genetic traits during the genetic process (such as reproduction, iteration, natural selection / algorithmic evolution). It can be obtained through statistical analysis of a large amount of genetic data, or through other methods, which will not be elaborated here.

[0059] Step 104: Call the pre-trained face image generation model to process the target genetic features and generate the face image of the target object.

[0060] Among them, the pre-trained face image generation model is a tool specifically designed to generate realistic face images based on given genetic features. During training, this model learns a large number of mapping relationships between face images and their corresponding genetic features, enabling it to transform target genetic features into visualized face images.

[0061] In this step, the target genetic features are input into a pre-trained face image generation model to generate a face image of the target object.

[0062] In this embodiment, a target kinship image of a blood relative of the target object is obtained; based on the target kinship image, the target kinship features corresponding to each structural unit are determined; a pre-trained feature analysis model is used to analyze the target kinship features corresponding to each structural unit to determine the target genetic features; a pre-trained face image generation model is invoked to process the target genetic features and generate a face image of the target object. This application can generate a face image of the target object based on the target kinship image. The generated face image can highly realistically reflect the current facial features of the maliciously transferred young object, providing crucial and powerful evidence for handling such cases.

[0063] To train the aforementioned feature analysis model so that it can accurately extract genetic features from kinship images, the embodiments of this application employ the following training steps: Step 1: Collect a large number of publicly available family photos or image data containing clear kinship relationships to construct a large-scale family kinship feature database, resulting in a sample dataset including family kinship relationships. Each sample in the sample dataset is a "family unit" containing at least one face image of the target object (children) and several face images of related relatives (such as father, mother, siblings, etc.).

[0064] Step 2: Using a pre-built feature extraction network (e.g., a deep convolutional neural network pre-trained on a large-scale face dataset or a publicly available face recognition model), process each image in each sample of the sample dataset. The feature extraction network acts as a feature encoder, mapping the input raw face image into a high-dimensional feature vector. This yields the baseline genetic feature vector corresponding to the target object in the sample, and the sample kinship feature vector corresponding to related objects.

[0065] Step 3: Construct a feature analysis model using the Transformer architecture. The model's input layer receives the sample kinship feature vectors corresponding to the aforementioned kinship objects. Utilizing the self-attention mechanism in the Transformer architecture, the model can capture the long-distance dependencies and association strengths between different sample kinship feature vectors, as well as between sample kinship feature vectors and potential genetic patterns.

[0066] Step 4: The feature analysis model also includes a weight predictor. The weight predictor connects to the Transformer module and, based on the contextual features output by the Transformer, predicts the contribution of each sample's kinship feature vector to the baseline genetic feature vector, outputting the corresponding feature weights. Subsequently, based on these feature weights, all sample kinship feature vectors are weighted and fused (e.g., weighted summation) to obtain the predicted genetic feature vector.

[0067] Step 5: During training, calculate the cosine similarity between the predicted genetic feature vector and the corresponding baseline genetic feature vector. Use the cosine similarity as the primary or sole metric of the loss function (e.g., loss = 1 - cosine similarity). Minimize this loss value using the backpropagation algorithm, continuously updating the parameters of the feature analysis model (including the Transformer module and weight predictor) until the model converges.

[0068] Through the above training method, the feature analysis model can learn how to accurately point to or approximate the true genetic characteristics of the target offspring in the vector space based on the combination of facial features of kinship objects, and finally obtain the trained feature analysis model.

[0069] In this embodiment, the face generation system has a publicly available face database. For each face image in the database, the following operations are performed to obtain the features corresponding to all structural units on each face. The specific steps include: processing the face image using a first deep neural network to obtain a feature map corresponding to the face image; simultaneously dividing the face image to obtain a set of face region masks, which includes the face region mask corresponding to each structural unit. Then, scaling the face region masks in the face region mask set to the same size as the feature map, and binarizing the scaled masks to obtain binarized features, and then binarizing the feature set. Multiplying the binarized features in the binarized feature set with the feature map to obtain a multiplication result, and inputting this multiplication result into a second deep neural network to obtain its corresponding features. That is, obtaining the features corresponding to each structural unit on the face image. Then, for multiple face images, combining the features of the same structural unit to obtain a feature set corresponding to each structural unit. For each feature set corresponding to a structural unit, using a clustering method to divide the features in the feature set to obtain multiple clustering results. Each clustering result essentially corresponds to a category. Therefore, the central feature in the clustering result is determined as the discrete feature corresponding to that category. All discrete features corresponding to this structural unit are combined to obtain its corresponding discrete feature set.

[0070] Alternatively, for each structural unit, a sub-image corresponding to each category can be obtained, and then the sub-image can be input into a feature extraction model to obtain the discrete features corresponding to each category. Then, the discrete features corresponding to all categories can be combined to obtain the discrete feature set corresponding to the structural unit.

[0071] Finally, the discrete feature set corresponding to each structural unit is used to form a facial genetic feature library.

[0072] It should be noted that the reason why this application uses the methods of 'discrete feature set' and 'discretization encoding' to determine the target genetic features is based on the inventor's profound insight into the problem of ambiguity in genetic features in the prior art.

[0073] In traditional deep learning-based face generation or style transfer techniques, facial features are typically mapped to a continuous high-dimensional vector space. When the model attempts to predict offspring features, it often employs weighted interpolation or regression of the parental feature vectors. However, this continuous space regression operation is prone to 'feature averaging' when dealing with highly nonlinear genetic features. For example, when the parents have 'high nose bridge' and 'flat nose bridge' respectively, the continuous vector model tends to generate a 'fuzzy intermediate state' feature between the two. This not only loses high-frequency texture details of the face, resulting in a distorted and overly smoothed facial image due to texture loss, but also fails to accurately reflect the discontinuous selection process of 'dominant / recessive' traits in genetics.

[0074] In contrast, the discretization encoding mechanism introduced in this application forces the predicted intermediate features to be mapped to a pre-constructed 'discrete feature set' consisting of clear and independent typical features. This is equivalent to introducing a 'quantization correction' process in the feature space: regardless of the intermediate results predicted by the model, the final determined target genetic feature must be a 'prototype feature' of a certain class that truly exists in the feature library and has clear texture. This mechanism effectively blocks the propagation of feature ambiguity, ensuring that each structural unit (such as eyes and nose) of the generated offspring face has highly realistic texture details and a clear morphological structure, thereby significantly improving the realism of face generation and the interpretability of genetic prediction.

[0075] In this embodiment, a pre-trained feature analysis model is first invoked to analyze the target kinship features corresponding to each structural unit, obtaining a first feature. Then, based on the first feature, discrete features are determined from the discrete feature set through similarity calculation. Finally, the target genetic feature is determined based on these discrete features. Therefore, this application provides a method for determining target genetic features, as follows: Figure 2 As shown, the specific steps include: Step 201: Call the pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and obtain the first feature corresponding to each structural unit.

[0076] Among them, the pre-trained feature analysis model is built on a large amount of data and professional algorithms. Its purpose is to deeply analyze the target kinship features corresponding to each structural unit, and to mine more representative and potentially valuable information, so as to obtain the first feature corresponding to each structural unit.

[0077] In this step, during the training phase, the feature analysis model learns from a massive amount of sample data containing various kinship characteristics. This data covers rich information from different families and individuals. Through complex machine learning algorithms, such as neural networks in deep learning, the model gradually grasps the intrinsic connections and patterns between various kinship characteristics. Then, the target kinship characteristic corresponding to each structural unit is input into the pre-trained feature analysis model to generate the first feature corresponding to each structural unit.

[0078] Step 202: For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the first feature corresponding to the structural unit.

[0079] In this step, for each structural unit, the discrete feature set corresponding to the structural unit is obtained, and a similarity calculation algorithm is used to calculate the similarity between each discrete feature in the discrete feature set and the first feature. The discrete features whose similarity satisfies the first preset condition are determined as the discrete features corresponding to the structural unit.

[0080] Step 203: Determine the target genetic features based on the discrete features corresponding to each structural unit.

[0081] In this step, the discrete features corresponding to all structural units can be fused to obtain the target genetic feature. The fusion method can be simple splicing or other fusion methods; it is not limited here.

[0082] In this embodiment, to improve data processing efficiency, discrete features are determined from a discrete feature set based on the target kinship features corresponding to the structural unit; then, a pre-trained feature analysis model is invoked to determine the target genetic features based on these discrete features. Therefore, this embodiment provides a method for determining target genetic features, the specific steps of which include: for each structural unit, obtaining a discrete feature set corresponding to the structural unit; determining the discrete features corresponding to the structural unit from the discrete feature set based on the target kinship features corresponding to the structural unit; and invoking a pre-trained feature analysis model to process the discrete features corresponding to each structural unit to determine the target genetic features.

[0083] In this step, there is a one-to-one correspondence between structural units and discrete feature sets. Therefore, for each structural unit, the corresponding discrete feature set is obtained, and the discrete feature with the highest similarity to the corresponding target kinship feature is found within this set. Finally, the discrete features corresponding to each structural unit are input into a pre-trained feature analysis model, which processes the input data to obtain the target genetic features.

[0084] In this embodiment, a preset feature analysis model is invoked to analyze the target kinship features of each structural unit, resulting in a corresponding probability set. For each structural unit, the category corresponding to the highest probability in the probability set is identified, and the discrete features corresponding to that category are used as the discrete features of this structural unit. Finally, based on the discrete features of all structural units, the target genetic features are determined. Therefore, this embodiment provides a method for determining target genetic features, as follows: Figure 3 As shown, the specific steps include: Step 301: Call the pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and obtain the probability set corresponding to each structural unit.

[0085] The feature analysis model learns the relationships between various genetic patterns and features during training, assigning a probability value to each possible category (e.g., eye shape categories like round or almond-shaped; eyelid features categories like single or double eyelids). These probability values ​​reflect the likelihood that the structural unit belongs to each category based on the current target's kinship features. The probability set includes the probability corresponding to each category.

[0086] In this step, the target kinship features corresponding to each structural unit are input into a pre-trained feature analysis model to obtain the probability set corresponding to each structural unit.

[0087] Step 302: For each structural unit, determine the category corresponding to the highest probability in the corresponding probability set, and determine the discrete features corresponding to the category as the discrete features corresponding to the structural unit.

[0088] Step 303: Determine the target genetic features based on the discrete features corresponding to each structural unit.

[0089] In this step, the discrete features corresponding to these structural units can be concatenated, and the concatenated features can be used as the target genetic feature. Alternatively, a feature fusion algorithm can be used to fuse the discrete features corresponding to these structural units to obtain the target genetic feature. Other methods can also be used; this is not limited to these methods.

[0090] In this embodiment, the target kinship characteristics corresponding to each structural unit are analyzed using a sub-attention module to obtain a second feature. This second feature is then processed using a feature processing module to finally derive the target genetic characteristic. Therefore, this embodiment provides a method for determining target genetic characteristics, as follows: Figure 4 As shown, the specific steps include: Step 401: Using the sub-attention module, analyze the target kinship features corresponding to each structural unit to obtain the second feature.

[0091] The sub-attention module focuses on key components of the target kinship features. Similar to human attention mechanisms, it assigns different weights to different features, emphasizing information closely related to genetic characteristics to derive a second feature. For example, when analyzing the target kinship features of a facial structural unit, the sub-attention module can focus on analyzing features such as shape and texture that significantly influence genetic traits within that unit, extracting a more representative second feature.

[0092] In this step, the target kinship features corresponding to each structural unit are input into the sub-attention module to obtain the second feature.

[0093] Step 402: Use the feature processing module to process the second feature to obtain the target genetic feature.

[0094] In this step, the feature processing module is used to process the second feature. Based on preset algorithms and rules, the feature processing module integrates, optimizes, and deeply mines the second feature, fusing and analyzing the second features of each structural unit to ultimately obtain the target genetic feature. For example, by performing weighted calculations and feature combinations on the second features of multiple structural units, a target genetic feature that comprehensively reflects the genetic characteristics of the target object can be generated.

[0095] In this embodiment of the application, in order to accurately generate a face image, the application can also obtain the age and gender information of the target object. Then, the target genetic features, age information, and gender information are input into a pre-trained face image generation model to generate a face image of the target object. Specific steps include: obtaining the attribute information of the target object; calling the pre-trained face image generation model to process the target genetic features and attribute information to generate a face image of the target object.

[0096] The attribute information includes age, gender, and other information.

[0097] In this step, a pre-trained face image generation model is invoked. During its training phase, this model utilized a massive amount of rich face image data, comprehensively covering face samples from different age groups, genders, and various combinations of genetic features. Through continuous optimization using deep learning algorithms, the model deeply learns and masters the complex mapping relationship between target genetic features, attribute information, and face images. In the usage phase, the acquired target genetic features and attribute information are input into the model, and the model gradually generates face images of the target object based on the learned mapping relationship.

[0098] In this embodiment, when facial features include sub-features corresponding to each structural unit, refined analysis of facial features is achieved by dividing the parent facial image and extracting the sub-features corresponding to each structural unit. This approach allows the understanding of facial features to extend beyond the overall picture to delve into each specific structural unit. Furthermore, in practical applications, sub-feature extraction helps improve the accuracy and reliability of facial recognition because sub-features of different structural units can provide richer recognition information. Specific steps include: dividing the target kinship image to obtain unit images corresponding to each structural unit; for each structural unit, obtaining the feature extraction model corresponding to the structural unit; and using the feature extraction model to process the unit images corresponding to the structural unit to obtain the target kinship features corresponding to the structural unit.

[0099] In this step, a specialized image segmentation algorithm is used to divide the target kinship image. For example, a deep learning-based semantic segmentation algorithm can learn the feature patterns of different structural units (such as eyes, nose, mouth, cheeks, etc.) in a facial image, thereby accurately segmenting the target kinship image into unit images corresponding to each structural unit. For each structural unit, the feature extraction module corresponding to that structural unit is called, and the sub-image corresponding to that structural unit is processed using that feature extraction module to obtain the sub-features corresponding to that structural unit.

[0100] During the training process, the aforementioned feature extraction module used a large amount of image data containing different facial structural units. Through deep learning algorithms, it learned the mapping relationship between sub-images and sub-features of different structural units.

[0101] Furthermore, to ensure more accurate feature extraction, target kinship features corresponding to each structural unit can be obtained based on the feature maps of the sub-images of each structural unit and the face image. A face segmentation algorithm model based on a deep neural network is then used to segment the target kinship images, generating mask images for each structural unit. The feature maps of the face image and the mask images of each structural unit are then input into a feature extraction model to extract the target kinship features for each structural unit. Specifically, the first step is to extract the feature map of the face image using a deep neural network; the second step is to scale the face region mask to the same size as the feature map and perform binarization; the third step is to multiply each mask image with the feature map to obtain a region feature map; and the fourth step is to feed the region feature map into the feature extraction model to obtain the target kinship features for each structural unit.

[0102] In this embodiment of the application, after obtaining the face image of the target object, the target object is searched in the public face database based on the face image. The specific steps are as follows: using a similarity calculation algorithm, the similarity between each face in the public face database and the newly generated adult face image is calculated, and a preset number of faces with the highest similarity are taken as faces similar to the newly generated adult face image, and the object corresponding to the face image is determined as the target object.

[0103] Furthermore, when multiple target objects are identified, the household registration information of each target object is obtained; for each target object, the corresponding household registration score is determined based on the corresponding household registration information; and based on the corresponding household registration score of each target object, the target objects that meet the preset conditions are determined among all target objects.

[0104] The higher the household registration score, the better the public security in the corresponding area, and the lower the probability of related adverse events. Conversely, the lower the household registration score, the worse the public security in the corresponding area, and the higher the probability of related adverse events. For example, if the target's household registration information is Beijing, its corresponding household registration score can be set to 9 because Beijing has good public security. If the target's household registration information is XX Province, XX City, XX County, XX Village, its corresponding household registration score can be set to 4 because this village is located in the mountains and many related incidents have occurred there. The preset conditions can be objects with household registration scores lower than a preset score or a preset number of objects with the lowest household registration scores. Of course, other preset conditions are also possible and are not limited here.

[0105] In this step, after identifying the target objects, for each target object, the household registration information of the target object is queried. Based on the pre-set correspondence between household registration information and household registration scores, the corresponding household registration score is determined and set as the household registration score for that target object. Finally, based on the household registration score corresponding to each target object, the target objects that meet the preset conditions are identified from all target objects.

[0106] In this embodiment, the target object can also be verified to determine if it is more likely to be the object being searched. Specific steps include: obtaining a preset verification strategy; and using the preset verification strategy to verify the target object.

[0107] In this step, a preset verification strategy can generate a verification notification for the system and send it to the relevant department in the target's household registration office, enabling the department to verify the target's information. For example, the verification notification could involve collecting DNA samples from the target and their current family members to determine if a blood relationship exists between them.

[0108] Based on the above embodiments, after obtaining a facial image, since the facial image excludes objective attribute differences such as gender and age, searching for targets based on this facial image can more accurately and efficiently locate individuals matching the facial features, reducing interference from non-core facial feature differences such as gender and age, improving the accuracy and targeting of the search, and finding the target object faster. Specific solutions include: Case 1: During the initial facial comparison of suspected individual A (out of 800 people), over 100 individuals with "similar appearances" were manually selected. However, due to A's young age at the time of disappearance and significant changes in appearance as an adult (age interference), even though A was on the candidate list, he was deemed "low probability" and shelved. Subsequently, the team used cross-age kinship comparison technology to eliminate age interference and re-identified suspected individual A. Ultimately, after two DNA verifications, it was confirmed that the man was indeed suspected individual A, who had been missing for 22 years.

[0109] Case 2: Suspect B was transferred to Region 1 at the age of 3. After reaching adulthood, he returned to Region 2 with his adoptive parents, while his biological parents remained in Region 2. Due to "differences in appearance between childhood and adulthood (age interference)," the two groups repeatedly missed each other without being identified. In 2023, the authorities, using facial recognition technology that excludes age- and gender-irrelevant features, successfully matched the adult face of suspect B with the face of the child who went missing, ultimately leading to their reunion on September 26.

[0110] Case 3: From 1993 to 1996, suspect XX and his accomplice XXX traveled to various locations, renting houses and familiarizing themselves with the local environment to target suitable young children for malicious relocation. They repeatedly used these methods to maliciously relocate 11 young children, including suspected individuals C and D, to region 3. Furthermore, of the 13 people suspected of being maliciously relocated by suspect XX, 10 have been reunited with their families of origin, 2 have been located by their suspected relatives, and 1 remains missing.

[0111] Case 4: Suspect F was maliciously moved when he was only 100 days old, and his parents searched for him for 25 years without success. After the authorities apprehended the suspect in 2022, they only learned that suspect F had been maliciously moved to a location near region 4. However, due to the large time span and the fact that suspect F's appearance had completely changed beyond infancy (age interference), the search reached a stalemate. In November 2023, the authorities used cross-age kinship comparison technology (using the core facial features of suspect F's parents as anchor points to eliminate interference from suspect F's age and gender) to quickly identify the suspect. After DNA confirmation, the parents were reunited on December 1st.

[0112] Beyond specific case studies, various data and charts also visually demonstrate the powerful effectiveness of the technology. Table 1 presents the number of matches using only early childhood characteristics and combined with same-kinship characteristics for different ranges of "topk" candidate numbers, clearly showing that same-kinship comparison can effectively improve the match rate, specifically: Table 1

[0113] Figure 5 The bar chart shows a significant increase in the number of cases involving the malicious transfer of minors solved annually, which indirectly confirms that the relevant technology has been used to find the information, and the images generated are more accurate. Figure 6 The bar chart shows the number of times the ranking of the 2 million database improved under different original database ranking ranges, reflecting the effect of technology on improving the ranking of search objects.

[0114] Meanwhile, looking at historical cases: 1. Suspect B: Traditional methods took 2 years without results; using the same kinship method, the algorithm results were obtained and analyzed in less than 1 hour; 2. Xiao C: less than 1 hour; 3. Xiao Xie: less than 1 hour; and some results were obtained in half an hour.

[0115] It should be understood that although the steps in the flowchart are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order constraint on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the diagram may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0116] Please see Figure 7 One embodiment of this application provides a face image generation apparatus 700, comprising: Acquisition unit 701 is used to acquire target kinship images of objects that are related to the target object by blood; The first determining unit 702 is used to determine the target kinship features corresponding to each structural unit based on the target kinship image; The second determining unit 703 is used to analyze the target kinship features corresponding to each structural unit using a pre-trained feature analysis model to determine the target genetic features; Calling unit 704 is used to call a pre-trained face image generation model to process the target genetic features and generate a face image of the target object.

[0117] Optionally, the second determining unit 703 is used for: The pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, and the first feature corresponding to each structural unit is obtained. For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the first feature corresponding to the structural unit; The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0118] Optionally, the second determining unit 703 is used for: For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the target kinship feature corresponding to the structural unit; A pre-trained feature analysis model is invoked to process the discrete features corresponding to each structural unit, thereby determining the target genetic feature.

[0119] Optionally, the second determining unit 703 is used for: A pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, thereby obtaining a probability set corresponding to each structural unit, wherein the probability set includes the probability corresponding to each category; For each structural unit, determine the category corresponding to the highest probability in the corresponding probability set, and determine the discrete features corresponding to the structural unit by using the discrete features corresponding to the category. The target genetic feature is determined based on the discrete features corresponding to each structural unit.

[0120] Optionally, the pre-trained feature analysis model includes a sub-attention module and a feature processing module, and the second determining unit 703 is used for: Using the sub-attention module, the target affinity features corresponding to each structural unit are analyzed to obtain the second feature; The second feature is processed using the feature processing module to obtain the target genetic feature.

[0121] Optionally, the calling unit 704 is used for: Obtain the attribute information of the target object; A pre-trained face image generation model is invoked to process the target genetic features and attribute information to generate a face image of the target object.

[0122] Optionally, the first determining unit 702 is used for: The target kinship image is divided to obtain a unit image corresponding to each structural unit; For each structural unit, the feature extraction model corresponding to the structural unit is obtained. The feature extraction model is used to process the unit image corresponding to the structural unit to obtain the target kinship feature corresponding to the structural unit.

[0123] For specific limitations regarding the aforementioned face image generation device, please refer to the limitations on the face image generation method described above, which will not be repeated here. Each unit in the aforementioned face image generation device can be implemented entirely or partially through software, hardware, or a combination thereof. Each of these units can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0124] In one embodiment, a computer device is provided, the internal structure of which can be as shown in the figure. Figure 8 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores data. The network interface communicates with external terminals via a network connection. The computer program, executed by the processor, can implement the above-described face image generation method. It includes: memory and a processor; the memory stores a computer program; and the processor executes the computer program to implement any step of the above-described face image generation method.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, can perform any of the steps in the face image generation method described above.

[0126] Those skilled in the art will understand that embodiments of this application can provide a face image generation method, system, or computer program product. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented 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.

[0127] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] 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.

[0129] 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.

[0130] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0131] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for generating a human face image, characterized in that, include: Obtain the target kinship image of objects that are related to the target object by blood; Based on the target kinship image, determine the target kinship features corresponding to each structural unit; Using a pre-trained feature analysis model, the target kinship features corresponding to each structural unit are analyzed to determine the target genetic features. This includes: calling the pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit, obtaining a first feature corresponding to each structural unit; for each structural unit, obtaining a discrete feature set corresponding to the structural unit; determining the discrete features corresponding to the structural unit from the discrete feature set based on the first feature corresponding to the structural unit; and determining the target genetic features based on the discrete features corresponding to each structural unit. A pre-trained face image generation model is invoked to process the target genetic features and generate a face image of the target object.

2. The method according to claim 1, characterized in that, The method uses a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features, including: For each structural unit, obtain the discrete feature set corresponding to the structural unit, and determine the discrete feature corresponding to the structural unit in the discrete feature set according to the target kinship feature corresponding to the structural unit; A pre-trained feature analysis model is invoked to process the discrete features corresponding to each structural unit, thereby determining the target genetic feature.

3. The method according to claim 1, characterized in that, The method uses a pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit and determine the target genetic features, including: A pre-trained feature analysis model is invoked to analyze the target kinship features corresponding to each structural unit, thereby obtaining a probability set corresponding to each structural unit, wherein the probability set includes the probability corresponding to each category; For each structural unit, determine the category corresponding to the highest probability in the corresponding probability set, and determine the discrete features corresponding to the structural unit by using the discrete features corresponding to the category. The target genetic feature is determined based on the discrete features corresponding to each structural unit.

4. The method according to claim 1, characterized in that, The pre-trained feature analysis model includes a sub-attention module and a feature processing module. The pre-trained feature analysis model is used to analyze the target kinship features corresponding to each structural unit to determine the target genetic features, including: Using the sub-attention module, the target affinity features corresponding to each structural unit are analyzed to obtain the second feature; The second feature is processed using the feature processing module to obtain the target genetic feature.

5. The method according to claim 1, characterized in that, The step of calling a pre-trained face image generation model to process the target genetic features and generate a face image of the target object includes: Obtain the attribute information of the target object; A pre-trained face image generation model is invoked to process the target genetic features and attribute information to generate a face image of the target object.

6. The method according to claim 1, characterized in that, The step of determining the target kinship features corresponding to each structural unit based on the target kinship image includes: The target kinship image is divided to obtain a unit image corresponding to each structural unit; For each structural unit, the feature extraction model corresponding to the structural unit is obtained. The feature extraction model is used to process the unit image corresponding to the structural unit to obtain the target kinship feature corresponding to the structural unit.

7. The method according to claim 1, characterized in that, The method further includes: Obtain the target description text; When the target description feature is not extracted from the target description text, the feature analysis model is invoked to process the target kinship feature to obtain the target genetic feature of the target object; the target description feature is the facial description feature of the target object; When the target description feature is extracted from the target description text, it is detected whether the target description feature includes genetic description feature and face description feature, and the feature in the detection result is obtained. The genetic feature analysis module is invoked to process the target kinship features and the features in the detection results to obtain the face image of the target object.

8. A face image generation device, characterized in that, The device includes: The acquisition unit is used to acquire target kinship images of objects that are related to the target object by blood. The first determining unit is used to determine the target kinship features corresponding to each structural unit based on the target kinship image; The second determining unit is used to analyze the target kinship features corresponding to each structural unit using a pre-trained feature analysis model to determine the target genetic features. This includes: calling the pre-trained feature analysis model to analyze the target kinship features corresponding to each structural unit, obtaining a first feature corresponding to each structural unit; for each structural unit, obtaining a discrete feature set corresponding to the structural unit; determining the discrete feature corresponding to the structural unit from the discrete feature set based on the first feature corresponding to the structural unit; and determining the target genetic features based on the discrete features corresponding to each structural unit. The calling unit is used to call a pre-trained face image generation model to process the target genetic features and generate a face image of the target object.

9. A computer device, comprising: A memory and a processor, the memory storing a computer program, characterized in that the processor, when executing the computer program, implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images