Face generation method and system based on reverse correlation

Abstract

The invention belongs to the technical field of psychology, and discloses a face generation method and system based on reverse correlation, and the method comprises the steps: generating a random noise image through employing a sine wave superposition method; using a face discrimination model to discriminate the random noisy point image, and judging whether face features exist or not; superposing the random noisy point images with the face features to obtain a single simulated face; and repeating the steps to obtain a plurality of simulated faces. The method provided by the invention solves the problems of complex face collection algorithm and low accuracy and diversity in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of psychology, and in particular to a face generation method and system based on reverse correlation. Background Art

[0002] Human faces have rich features, which are not only reflected in visual differences, but also in emotional expression, self-protection, social interaction and other aspects. Due to the importance and diversity of facial features, it is particularly important to obtain as many face samples as possible. This is not only to meet the needs of academic research, but also to meet the needs of practical applications. The study of facial features has important application value in many fields such as psychology, cognitive neuroscience and sociology: In psychological research, facial expressions are considered to be an important way to reveal the emotional state of individuals. Through the analysis of facial expressions, researchers can better understand the emotional mechanism of human beings. For example, emotions such as anger, sadness, and happiness can be expressed through facial expressions, and these emotional reactions will have different manifestations in different situations. These studies not only contribute to the development of psychological and cognitive neuroscience theories, but can also be applied to actual psychological counseling and treatment to help individuals better regulate their emotions. In social interactions, facial expressions and features are important tools for people to communicate non-verbally. Through facial expressions, people can convey complex social signals such as agreement, rejection, suspicion, etc. These signals play a vital role in daily communication. For example, in business negotiations, facial expressions can convey the true intentions and attitudes of negotiators, thereby affecting the outcome of negotiations. In interpersonal relationships, facial expressions are also an important means of expressing emotions and building trust. Therefore, the study of facial features is of great significance for understanding social behavior and improving interpersonal relationships. In terms of self-protection, facial features can be used to identify individual identities, thereby ensuring personal safety. For example, face recognition technology has been widely used in various security systems, such as access control systems and mobile phone unlocking. These technologies confirm individual identities by analyzing facial features, effectively preventing identity theft and illegal intrusion. At the same time, the study of facial features can also help improve the accuracy and reliability of these technologies, further ensuring the safety of individuals and society.

[0003] At present, common face collection methods mainly include collecting real human faces and using algorithms to simulate and generate faces. However, these methods all have certain problems and challenges in practical applications. First of all, collecting real human faces requires a large amount of image data, which usually involves complex ethical issues and high costs. During the collection process, researchers need to ensure the informed consent of the subjects, protect their privacy, and comply with relevant laws and regulations. This not only increases the difficulty and cost of data collection, but may also limit the diversity and representativeness of the data. Especially when it comes to sensitive groups or special populations, the ethical issues of data collection are more prominent.

[0004] In addition, the collection of real face data is also limited by the facial styles in the database, which may not cover the facial features of all races, age groups and genders, resulting in insufficient representation of the data. Since facial features vary significantly between different races, ages and genders, a lack of diverse databases may lead to biased research results.

[0005] In order to alleviate the above problems, methods of using algorithms to simulate and generate faces have emerged in recent years. These methods generate virtual faces through computer algorithms, which can not only save the cost of data collection, but also solve ethical issues to a certain extent. However, algorithm-generated faces also have limitations. First, these generated faces usually rely on existing training data. If the training data is not diverse enough, the generated faces will also lack diversity. In addition, simulated faces may not be as realistic as real faces in details, especially in expressing subtle emotions and individual characteristics. These problems limit the scope of application and effect of simulated faces.

[0006] Existing technologies mainly focus on simulating real faces as much as possible, but there is still much room for improvement in terms of diversity and detail. For example, existing generation algorithms may have deviations when representing facial features of different races, and the generated faces may not accurately reflect facial features in different cultural backgrounds. In addition, existing generation technologies are also insufficient in expressing emotional expressions and individual unique characteristics, and the generated faces may lack realism and subtlety. These problems limit the application of simulated generated faces in fields such as psychology research, social interaction analysis, and human-computer interaction.

[0007] In summary, the rich features of human faces have important application value in many fields. However, existing face collection and generation methods have many challenges and limitations in terms of data diversity, ethical issues, cost, and detail expression. Solving these problems requires further research and improvement in technology and methods to improve the representativeness, authenticity, and application effect of face data. Summary of the invention

[0008] The embodiments of the present application provide a face generation method and system based on reverse correlation to at least solve the technical problems listed in the prior art.

[0009] A face generation method based on reverse correlation, wherein the method comprises the following steps:

[0010] S10. Generate a random noise image;

[0011] S20. using a face recognition model to recognize random noise images to determine whether there are facial features;

[0012] S30. Superimposing random noise images with facial features to obtain a single simulated face;

[0013] S40. Repeat step S30 to obtain multiple simulated faces.

[0014] Furthermore,

[0015] Step S10 includes generating the random noise image by using a sine wave superposition method.

[0016] Furthermore, step S10 further includes:

[0017] S101. Determine the size of the random noise image, generate a matrix of a specified size, and arrange the elements in the matrix in an increasing order;

[0018] S102. Rotate the matrix generated in S101 according to the set rotation angle to generate a rotation matrix;

[0019] S103. Introducing a phase offset, adding the phase offset to the rotation matrix generated in S102, performing a sine operation to generate a noise unit image;

[0020] S104. Multiply the elements in the noise unit image by the random matrix one by one to generate a random noise point image.

[0021] Furthermore,

[0022] In S101 , a matrix of a specified size is generated. The matrix sizes include: 16×16 and 32×32.

[0023] Furthermore,

[0024] S102 generates a rotation matrix including the following steps:

[0025] First set the rotation angle R, its value: R = {0, 30, 60, 90, 120, 150};

[0026] Then substitute the rotation angle R into the following formula to calculate the matrix S r :

[0027]

[0028] Furthermore,

[0029] In step S103, the phase offset is first set Then the phase shift P is substituted into the following formula to calculate the noise unit image I unit :I unit = sin(S r +p),p∈P.

[0030] Furthermore,

[0031] The random matrix value range in step S104 is between -1 and 1.

[0032] Furthermore,

[0033] Step S20 specifically includes:

[0034] S201. Using a neural network to extract image features of the random noise image generated in step S10;

[0035] S202. Using a face recognition model to determine the image features, the probability of a face existing in the random noise image is determined;

[0036] Furthermore, the method further includes the step of classifying the image features extracted in step S201 using a face classification model.

[0037] Furthermore, step S30 also includes: iterating and marking a plurality of noise images respectively, thereby obtaining a set of noise images covering faces and a set of noise images not covering faces, and combining the noise images in the two image sets to obtain a simulated face image.

[0038] The present invention also discloses a face generation system based on reverse correlation, the system comprising:

[0039] Noise generation module, used to generate random noise images;

[0040] A discrimination module is used to use a face discrimination model to discriminate random noise images to determine whether there are facial features;

[0041] The superposition module is used to superimpose random noise images with facial features to obtain simulated faces.

[0042] Furthermore, the noise generation module is used to generate the random noise image by using a sine wave superposition method. The noise generation module determines the size of the random noise image, generates a matrix of a specified size, and the elements in the matrix are arranged in an increasing order; then the generated matrix is ​​rotated according to a set rotation angle to generate a rotation matrix; then a phase offset is introduced, and the generated rotation matrix is ​​added with the phase offset, and a sine operation is performed to generate a noise unit image; finally, the elements in the noise unit image are multiplied one by one with the random matrix to generate a random noise image.

[0043] Furthermore, the discrimination module is used to use a neural network to extract image features of the random noise image generated by the noise generation module; then use a face discrimination model to judge the image features and judge the probability of the presence of a face in the random noise image; and classify the extracted image features.

[0044] Furthermore, the superposition module is used to iterate and mark a number of noise images respectively, thereby obtaining a set of noise images covering faces and a set of noise images not covering faces, and combining the noise images in the two image sets to obtain a simulated face image.

[0045] Compared with the prior art, the beneficial effects of the present invention include:

[0046] 1. Reduce the dependence on training data: Traditional face generation methods usually require a large amount of training data to ensure the quality of generation. The present invention can maintain a high generation quality with relatively less training data through reverse correlation technology, reducing the difficulty and cost of data acquisition. The reverse correlation: represents a technology that reveals the relationship between images and observer responses through statistical methods in image classification. This method involves controlled changes in random noise in images to estimate the relative influence of different noise patterns on human responses. In this way, intuitive images can be generated that reflect the way humans process visual information. In actual operation, the reverse correlation technology is implemented by adding specific random noise to the image and recording human responses to these changed images. Subsequently, statistical methods are used to analyze how different noise patterns affect human visual perception, thereby obtaining the weight of each noise pattern. Ultimately, these weights are used to generate a new image that is more visually induced by a specific human response.

[0047] 2. Enhance the stability and robustness of the generated model: The reverse correlation method can effectively utilize the noise and unstable factors that appear in the generation process, improve the stability and robustness of the generated model, and ensure that rich faces can be generated under the same parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0049] Figure 1 is a flow chart of a face generation method based on reverse correlation according to an embodiment of the present invention.

[0050] Figure 2 4 is a schematic diagram of the structure of a face generation system based on reverse correlation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0052] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0053] like Figure 1 As shown, a face generation method based on reverse correlation according to an embodiment of the present invention includes the following steps:

[0054] S10, using a sine wave superposition method to generate a random noise image.

[0055] The sine wave superposition method is as follows: first, after the required image size is determined, a matrix of the specified size is generated, and each element is arranged according to a specific increasing rule; then a set of rotation angles is set, and the corresponding rotation matrix is ​​generated by performing cosine and sine operations on the matrix; then a set of phase offsets is introduced, and the rotation matrix is ​​added with these phase offsets and sine operations are performed to generate a noise unit image. Finally, a random matrix with a value range between -1 and 1 is generated, and it is multiplied element by element with the noise unit image to obtain an image with random noise.

[0056] In one embodiment, the above-mentioned method of generating a random noise image by using a sine wave superposition method specifically includes the following steps S101-S104:

[0057] S101, confirm the image size to be generated, for example, set the image size to 16×16. In this example, choose to generate a 16×16 image matrix, and generate a sequence matrix A with the same columns and increasing rows from 0 to 2. 16×16 :

[0058]

[0059] The above 16×16 image matrix is ​​a relatively small matrix that can be used to quickly verify the effectiveness of the generation process. To improve accuracy, a larger matrix size can be selected as needed, such as 32×32 or even larger.

[0060] S102, set the angle R to be rotated, the value of which can be: R = {0, 30, 60, 90, 120, 150}, and substitute it into the following formula to calculate S r .

[0061]

[0062] Specifically, the matrix S r The generation process is as follows: the matrix A 16×16 Multiply it by cos(r) and then transform the matrix A 16×16 The transposed matrix of Multiply it by sin(r), add the two components together, and multiply the result by 2π.

[0063] Where r belongs to a given range R. The choice of angle R can be adjusted according to specific needs. In this example, we choose six different rotation angles of 0, 30, 60, 90, 120 and 150 degrees.

[0064] S103, setting phase offset And use the following formula to calculate the noise unit image I unit .

[0065] Iunit = sin(S r +p),p∈P

[0066] Specifically, I unit The generation process is as follows: the matrix S r Add a phase shift p and perform a sine operation, where p belongs to a given position P.

[0067] S104, generate a random matrix R with a value range of [-1,1] 16×16 , and with I unit Multiply the matrix elements to get a random noise image.

[0068] Specifically, generate a random matrix R with a value range between -1 and 1 16×16 This random matrix is ​​obtained by generating a random number at each element position. The random number between -1 and 1 can ensure that the value of the noise image is distributed in a reasonable range and avoid excessively large values. Then, this random matrix R 16×16 With the noise unit image I unit Multiply element by element to get the final random noise image. This step is to introduce randomness into the generated noise image so that each generated noise image has different characteristics.

[0069] S20. Use the face recognition model and combine it with the random noise image to generate a simulated face.

[0070] Face recognition model: This model is implemented through machine learning or deep learning technology, and aims to determine whether there is a face in the input image. First, a large number of labeled face and non-face image datasets are collected. Then, the image is preprocessed and feature extracted, including grayscale, scale normalization, and edge detection. Deep learning methods can automatically extract features directly from the image. Then, a suitable algorithm such as support vector machine (SVM), convolutional neural network (CNN), or random forest is selected to build a classifier. SVM is suitable for small high-dimensional data sets, CNN performs well in large-scale data sets and image recognition, and random forest provides strong classification and noise resistance by integrating multiple decision trees. During the training process, the model continuously optimizes parameters through a large amount of training data to improve the recognition accuracy. Finally, the trained model can accurately determine whether there is a face in the input image.

[0071] In one embodiment, step S20 specifically includes the following steps S201-S202:

[0072] S201. Use the S10 method to obtain a random noise image, and use the ResNet deep neural network to extract image features.

[0073] Specifically, the feature vector of the image can be obtained by encoding the random noise image through the pre-trained ResNet deep neural network. First, the random noise image is input into the pre-trained ResNet model. The ResNet model has been trained on a large image dataset and therefore has good feature extraction capabilities. The random noise image is input into the pre-trained ResNet model. After being processed by multiple convolutional layers and residual blocks of ResNet, the noise image is converted into a high-dimensional feature vector. These feature vectors contain important information in the image, such as edges, textures, and shapes, which provide a basis for subsequent face recognition.

[0074] S202, using the image features obtained in S201, judging the probability of the existence of a face according to a face discrimination model, and discriminating the random noise image.

[0075] In one embodiment, the random noise image is classified using the image features obtained in step S201 according to a trained face classification model, for example, using a face classification SVM model to determine the probability p(I) of the random noise image, which represents the possibility that the random noise image covers a face. Subsequently, the random noise is negatively valued, and the image features obtained in step S201 are used to classify the image according to the trained face classification SVM model to obtain the probability p(-I) of a negative random noise image, which represents the possibility that the random noise image does not cover a face. If p(I)>p(-I), the random noise image is considered to cover a face and is marked as 1, otherwise it is marked as 0.

[0076] S30, obtaining random noise images with facial features, and superimposing them to obtain a single simulated face.

[0077] Specifically, a large number of noise images are randomly generated, and step S20 is repeated multiple times to process and classify each noise image. For each iteration, the label value (0 or 1) of each noise image is recorded. Through multiple iterations and label accumulation, a set of two types of random noise images can be obtained: one is an image marked as 1, that is, a noise image that is considered to cover faces; the other is an image marked as 0, that is, a noise image that is considered not to cover faces. As the number of iterations increases, these two types of images are averaged and normalized, respectively, and their pixel values ​​are normalized so that all pixel values ​​are in the same range. This step is to eliminate the numerical differences between different images and make the features of the image more prominent.

[0078] When the correlation coefficient of two types of images reaches a threshold (e.g. -0.9), it can be considered that the two types of images have a high negative correlation. At this time, superimposing these images can obtain an image with simulated facial features. Through the superposition operation, we can combine the noise image containing facial features with the noise image without facial features to obtain a new simulated facial image.

[0079] In this process, the reverse correlation technology significantly improves the ability of the generative model to generate faces other than training data by effectively utilizing random noise and unstable factors in the generation process. For example, a generative model trained with smiling faces can generate simulated faces with sad facial features. This method not only enables the model to maintain high-quality output under a variety of environmental conditions, but also ensures the consistency and repeatability of the model output under the same parameter configuration, thereby enhancing the robustness and stability of the generative model. Such characteristics are particularly suitable for practical applications, and can effectively deal with common anomalies and deviations in real-world data, improving the practicality and generalization ability of the model.

[0080] S40, repeating step S30 to obtain multiple simulated faces.

[0081] Specifically, by continuously generating new random noise images and repeating step S20, a large number of simulated face images can be obtained. These simulated face images are images with facial features extracted from a large number of random noise images by screening and marking, providing rich data resources for subsequent image processing and analysis. The reverse correlation technology adopted by the present invention significantly reduces the demand for large-scale labeled data in this process, because it can use limited training samples to optimize and adjust the quality of generated images through human visual response. This method not only reduces the difficulty and cost of data collection, but also maintains the accuracy of image generation, so that even in an environment with limited data resources, high-quality image classification and face recognition can be effectively performed, greatly expanding its scope of application and practicality.

[0082] like Figure 2 The present invention also provides a face generation system based on reverse correlation, which includes the following modules:

[0083] Noise generation module, used to generate random noise images;

[0084] A discrimination module is used to use a face discrimination model to discriminate random noise images to determine whether there are facial features;

[0085] The superposition module is used to superimpose random noise images with facial features to obtain simulated faces.

[0086] The face generation system based on reverse correlation of the present invention includes the above-mentioned noise generation module, discrimination module and superposition module. Its working principle, working process, execution steps of each module, functions of each module and technical effects achieved are the same as or corresponding to the method described in the embodiment of the face generation method based on reverse correlation of the present invention, and will not be repeated here.

[0087] The simulated face images generated by this system are images with facial features extracted from a large number of random noise images through screening and labeling, providing rich data resources for subsequent image processing and analysis.

[0088] The above embodiments of the present application are only for description and do not represent the advantages and disadvantages of the embodiments. In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0089] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A face generation method based on reverse correlation, characterized in that: The method comprises the following steps: S10. Generate a random noise image; S20. using a face recognition model to recognize random noise images to determine whether there are facial features; S30. Superimposing random noise images with facial features to obtain a single simulated face; S40. Repeat step S30 to obtain multiple simulated faces.

2. The face generation method according to claim 1, characterized in that: Step S10 includes generating the random noise image by using a sine wave superposition method.

3. The face generation method according to claim 1 or 2, characterized in that: Step S10 also includes: S101. Determine the size of the random noise image, generate a matrix of a specified size, and arrange the elements in the matrix in an increasing order; S102. Rotate the matrix generated in S101 according to the set rotation angle to generate a rotation matrix; S103. Introducing a phase offset, adding the phase offset to the rotation matrix generated in S102, performing a sine operation to generate a noise unit image; S104. Multiply the elements in the noise unit image by the random matrix one by one to generate a random noise point image.

4. The face generation method according to claim 3, characterized in that: In S101 , a matrix of a specified size is generated. The matrix sizes include: 16×16 and 32×32.

5. The face generation method according to claim 3, characterized in that: S102 generates a rotation matrix including the following steps: First set the rotation angle R, its value: R = {0, 30, 60, 90, 120, 150}; Then substitute the rotation angle R into the following formula to calculate the matrix S r :

6. The face generation method according to claim 3, characterized in that: In step S103, the phase offset is first set Then the phase shift P is substituted into the following formula to calculate the noise unit image I unit :I unit = sin(S r +p),p∈P.

7. The face generation method according to claim 3, characterized in that: The random matrix value range in step S104 is between -1 and 1.

8. The face generation method according to claim 3, characterized in that: Step S20 specifically includes: S201. Using a neural network to extract image features of the random noise image generated in step S10; S202. Using a face recognition model to determine the image features, the probability of a face existing in the random noise image is determined; The method also includes the step of classifying the image features extracted in step S201 using a face classification model.

9. The face generation method according to claim 3, characterized in that: Step S30 also includes: iterating and marking a plurality of noise images respectively, thereby obtaining a set of noise images covering faces and a set of noise images not covering faces, and combining the noise images in the two image sets to obtain a simulated face image.

10. A face generation system based on reverse correlation, characterized in that: The system includes: Noise generation module, used to generate random noise images; A discrimination module is used to use a face discrimination model to discriminate random noise images to determine whether there are facial features; The superposition module is used to superimpose random noise images with facial features to obtain simulated faces.

Images