Method and system for generating face multi-view image based on single view

Abstract

The invention discloses a method and system for generating a face multi-view image based on a single view, and relates to the technical field of image generation and processing, and the method comprises the steps: S1, extracting feature points based on a single input face image, and constructing a face three-dimensional model, S2, carrying out the view transformation of the face three-dimensional model, generating a multi-view face image, and S3, carrying out the multi-view face image. S3, optimizing the generated multi-view-angle images to reduce the phenomenon of inconsistent three-dimensional space structures caused by parallax, and S4, integrating the optimized multi-view-angle images to generate a final multi-view-angle image; according to the method and the system for generating the face multi-view image based on the single view, the problem of inconsistent three-dimensional space structure caused by overlarge parallax when the multi-view image is generated in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image generation and processing, and in particular to a method and system for generating multi-view images of human faces based on a single view. Background Art

[0002] In the field of image generation and processing technology, methods for generating multi-view images of faces based on a single view have received widespread attention in recent years. Such methods use a single-view face image to generate multiple images from different perspectives through computer vision and deep learning technology, and are widely used in virtual reality, video synthesis, identity recognition and other fields.

[0003] Existing technologies usually rely on models such as convolutional neural networks or generative adversarial networks. These methods can generate relatively realistic multi-view face images by learning a large amount of three-dimensional face data. However, when generating multi-view images, existing methods often lead to inconsistency in the three-dimensional spatial structure due to the influence of excessive parallax. Specifically, when the image perspective changes greatly, the geometric shape and detailed features of the face (such as the relative positions of the eyes, nose, and mouth) cannot be accurately maintained, thereby affecting the realism and continuity of the image. The main reason for this problem is that the existing algorithms lack accurate modeling of three-dimensional geometric structures or are insufficient in parallax compensation. In addition, in extreme viewing angles, obvious artifacts or structural distortions may appear in the generated images, further reducing the quality and application effect of the generated images. When solving these problems, existing methods usually rely on more complex model structures or a large amount of annotated data, which undoubtedly increases the computational cost and implementation difficulty of the algorithm, limiting its application in real-time scenarios. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for generating multi-view images of a face based on a single view, so as to solve the problem of inconsistent three-dimensional spatial structure caused by excessive parallax when generating multi-view images in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for generating multi-view images of a face based on a single view, the method comprising: S1, extracting feature points based on a single input face image and constructing a 3D face model; S2, performing view transformation on the three-dimensional face model to generate face images from multiple perspectives; S3, optimizing the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax, including taking the distorted area in the image as the optimization target, simulating local adjustment of the image, redistributing the pixel positions based on the significant feature points, and updating the optimized parallax-corrected image. The specific formula is: Δd=k×F; Wherein, Δd represents the pixel offset required to adjust the inconsistent three-dimensional spatial structure area caused by parallax, k represents the pixel response value to the correction strength, and F represents the strength value required to be applied when optimizing the target area; S4. Integrate the optimized multi-view images to generate a final multi-view image.

[0006] Preferably, S1 includes extracting facial feature points from the input image, fitting the feature points to generate a three-dimensional surface model, optimizing the parameters by the least squares method, and outputting a three-dimensional model generated based on the feature points. The specific formula is: ; Among them, z represents the depth value of the facial feature point, x and y represent the horizontal and vertical positions of the facial feature point in the two-dimensional plane image, a represents the influence of x on the depth z, b represents the influence of y on the depth z, and c represents the overall offset.

[0007] Preferably, S2 includes: The initial position of the 3D face model is set as the default perspective, and the model is rotated around different axes to generate the target perspective. The specific formula is: ; Among them, u, v, w represent the horizontal coordinate, vertical coordinate, and depth coordinate of a point in the three-dimensional face model, respectively; u′, v′, w′ represent the new horizontal position, vertical position, and depth position of the point in the three-dimensional space after the rotation transformation, respectively; and θ represents the angle of rotation around a certain axis.

[0008] Preferably, S4 includes: The weights are set for the multi-view images by region, and multiple images are fused by pixel to generate a full-view consistency image. The specific formula is: ; Where I represents the final value of a pixel of the target view image when integrating multi-view images, n represents the number of view images generated and involved in fusion, i represents the index number of the multi-view image, and h represents the index number of the multi-view image. i Indicates the weight of the i-th view in the fusion process, I i Represents the gray value of a pixel at a corresponding position in the i-th perspective image; Perform edge sharpening on the fused image.

[0009] Preferably, the value k of the pixel response degree to the correction force in S3 ranges from 0.1 to 1.0.

[0010] Preferably, the calculation formula of the intensity value F to be applied when optimizing the target area in S3 is: ; Wherein, F represents the intensity value to be applied when optimizing the target area, G represents the grayscale value of the pixel point in the target area, p represents the horizontal position of the pixel point in the target area, and q represents the vertical position of the pixel point in the target area.

[0011] Preferably, the facial feature points in S1 are extracted based on key feature areas of the face, including the corners of the eyes, the tip of the nose, the corners of the mouth and the chin area.

[0012] Preferably, the value range of the angle θ of rotation around a certain axis in S2 is [-180°, 180°], and image sequences of different target viewing angles are generated by adjusting the angle.

[0013] Preferably, the edge sharpening process of the fused image in S4 is implemented by a Laplace operator, and the specific formula is: ; Among them, L′ represents the brightness value of a pixel in the image after Laplace operator processing, L represents the brightness value of a pixel in the input image, and λ represents the sharpening coefficient. Represents the gradient operator.

[0014] A system for generating multi-view images of a human face based on a single view, used to implement the steps of the method for generating multi-view images of a human face based on a single view, the system comprising: A feature extraction module is used to extract feature points based on a single input face image and construct a 3D face model; A view transformation module connected to the feature extraction module is used to perform view transformation on the three-dimensional face model to generate face images from multiple perspectives; An image optimization module connected to the view transformation module is used to optimize the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax; The image integration module connected to the image optimization module is used to integrate the optimized multi-view images to generate a final multi-view image.

[0015] It can be seen from the above technical solution that the present invention has the following beneficial effects: The method and system for generating multi-view images of a face based on a single view extract feature points based on a single input face image, construct a three-dimensional face model, perform view transformation on the three-dimensional face model, generate face images of multiple viewpoints, optimize the generated multi-view images, reduce the three-dimensional space structure inconsistency caused by parallax, integrate the optimized multi-view images, and generate the final multi-view images. The method and system can effectively reduce the three-dimensional space structure inconsistency caused by excessive parallax. The generated multi-view face images are more realistic and consistent in terms of the geometric shape and detail features of the face, and can significantly reduce artifacts and structural distortions in extreme viewpoints. The phenomenon improves the overall quality and visual effect of the generated image. While maintaining the high quality of the generated image, it avoids the dependence on complex model structure or a large amount of labeled data, reduces the computational cost and implementation difficulty of the algorithm, and can better adapt to the needs of real-time scenarios, such as virtual reality, video synthesis, and identity recognition. The high-quality multi-view images generated by the present invention have stronger continuity and authenticity, further improving the application effect in the fields of virtual reality, film and television production, game design, etc., meeting the needs of more scenarios, and solving the problem of inconsistent three-dimensional spatial structure caused by excessive parallax when generating multi-view images in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a schematic diagram of the connection of the system modules of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the present invention provides a technical solution: a method for generating a multi-view image of a face based on a single view, the method comprising: S1, extracting feature points based on a single input face image and constructing a 3D face model; S2, performing view transformation on the three-dimensional face model to generate face images from multiple perspectives; S3, optimizing the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax, including taking the distorted area in the image as the optimization target, simulating local adjustment of the image, redistributing the pixel positions based on the significant feature points, and updating the optimized parallax-corrected image. The specific formula is: Δd=k×F; Wherein, Δd represents the pixel offset required to adjust the inconsistent three-dimensional spatial structure area caused by parallax, k represents the pixel response value to the correction strength, and F represents the strength value required to be applied when optimizing the target area; S4. Integrate the optimized multi-view images to generate a final multi-view image.

[0019] In the above embodiment, the method is based on a single input face image, firstly uses a specific algorithm to extract face feature points (such as eyes, nose tip, mouth corners, etc.), and constructs a face three-dimensional model through the geometric relationship of the feature points. Based on this three-dimensional model, the view transformation technology is applied to generate face images of multiple different perspectives, such as left and right side, top view and top view. In order to ensure that the generated multi-perspective image is more realistic and consistent in visual effect, the three-dimensional space error area of ​​the generated image is further optimized, the pixel distribution is adjusted by local significant features, the influence of parallax is reduced, and the optimization calculation is performed by the formula Δd=k×F. Finally, the optimized multi-perspective images are integrated to output realistic multi-perspective face images. Through this method, the parallax problem existing in the traditional single-view face generation method can be effectively solved, so that the generated multi-perspective image has higher authenticity and consistency. Specifically, the present invention can make full use of feature point information when constructing a three-dimensional face model to ensure the accuracy of the three-dimensional model; in the multi-perspective generation stage, the local area distortion problem caused by parallax is optimized, and the overall visual effect of the generated image is improved. In addition, by precisely adjusting the pixel distribution, the generated images have significant improvements in spatial structural consistency and resolution, which is beneficial to practical applications in fields such as face recognition, virtual reality, and animation production.

[0020] S1 includes extracting facial feature points from the input image, fitting the feature points to generate a three-dimensional surface model, optimizing the parameters through the least squares method, and outputting a three-dimensional model generated based on the feature points. The specific formula is: ; Among them, z represents the depth value of the facial feature point, x and y represent the horizontal and vertical positions of the facial feature point in the two-dimensional plane image, a represents the influence of x on the depth z, b represents the influence of y on the depth z, and c represents the overall offset.

[0021] In the above implementation, the method of the present invention is based on the input face image, and the key points (such as eyes, nose tip and mouth corners) are accurately located by facial feature point extraction technology, and a three-dimensional surface model is generated based on this. Specifically, the coordinate data of the feature points are optimized and fitted by the least squares method to obtain a mathematical model that satisfies the depth information. The formula form of the model is , where a and b represent the weighted influence on the depth value in the horizontal and vertical directions, respectively. This optimization model can significantly reduce the impact of noise or feature point extraction errors on the accuracy of 3D surface modeling, and at the same time globally adjust the overall depth through the offset parameter c, thereby constructing a 3D model that is more in line with the real face geometry. Compared with the traditional face 3D modeling method, this method can effectively reduce the modeling error caused by noise or local feature point deviation by optimizing parameter fitting through the least squares method, thereby improving the accuracy and stability of the 3D model. Formula A simple but efficient linear relationship description is used, which is convenient for fast calculation and implementation. In addition, the optimized 3D model can better reflect the geometric structure characteristics of the face and enhance the realism of subsequent perspective transformation and image generation. This method is particularly suitable for scenarios that require real-time modeling, such as virtual reality and face recognition systems, and further improves modeling efficiency and robustness.

[0022] S2 includes setting the initial position of the 3D face model as the default perspective, rotating the model around different axes, and generating the target perspective. The specific formula is: ; Among them, u, v, w represent the horizontal coordinate, vertical coordinate, and depth coordinate of a point in the three-dimensional face model, respectively; u′, v′, w′ represent the new horizontal position, vertical position, and depth position of the point in the three-dimensional space after the rotation transformation, respectively; and θ represents the angle of rotation around a certain axis.

[0023] In the above implementation, the present invention performs a perspective transformation based on a rotation matrix by setting the initial position of the 3D face model to a default perspective. The rotation matrix defines a mathematical expression for rotation around a specific axis: For rotation around the vertical axis (Z axis), the two-dimensional rotation matrix part in the formula is applied to adjust the horizontal coordinate u and the vertical coordinate v so that the face model generates left and right side views from the default front view. By changing the rotation angle θ, different target view angles can be generated, θ>0 means clockwise rotation, and θ<0 means counterclockwise rotation. The linear operation of the rotation matrix ensures that the transformation of each point in the three-dimensional space is accurate and efficient. All feature points are redistributed to the target position after rotation transformation, thereby generating a multi-view face image. In this process, the rotation axis or the rotation angle range can be adjusted according to actual needs to generate different combinations of view angles. By using the view angle transformation method based on the rotation matrix, the present invention can accurately and quickly generate face images of multiple target view angles. Compared with the traditional image stretching or interpolation method, the method can better maintain the geometric consistency and spatial structure characteristics of the original image. The rotation matrix has a low computational complexity and can be used to generate multi-view images in real time. It is particularly suitable for application scenarios that require dynamic adjustment of the view angle, such as virtual reality, augmented reality, and multi-view face recognition. In addition, the method supports flexible adjustment of the view angle range and angle distribution, further improving the flexibility and applicability of the generated image.

[0024] S4 includes setting weights for multi-view images by region, fusing multiple images by pixel, and generating a full-view consistency image. The specific formula is: ; Where I represents the final value of a pixel of the target view image when integrating multi-view images, n represents the number of view images generated and involved in fusion, i represents the index number of the multi-view image, and h ... i Indicates the weight of the i-th view in the fusion process, I i Represents the grayscale value of a pixel at a corresponding position in the i-th viewing angle image; performs edge sharpening on the fused image.

[0025] In the above implementation, the present invention sets weights for multiple view images by region, and performs weighted averaging on the pixel values ​​at the same position based on the weights, thereby fusing and generating a full-view consistency image. Formula ; Describes the specific process of pixel fusion: Weight setting: weight h i Dynamically adjust the image according to the saliency of the viewpoint or its proximity to the target viewpoint, for example, assigning a higher weight to images closer to the target viewpoint to ensure visual consistency and detail fidelity of the fused image; Pixel fusion: Generate a new pixel value I by performing weighted average calculation on all participating view images at each pixel position. This method effectively eliminates the splicing traces between the boundaries of multi-view images and enhances the overall coherence of the image; Edge sharpening: In order to further improve the visual effect of the image, the fused image is sharpened to enhance edge details and make the image clearer and more natural. Common edge sharpening methods include Laplace filtering, Sobel operator, etc.

[0026] Compared with the simple image stitching method, the full-view consistency image generated by the weighted fusion method of the present invention has higher smoothness and authenticity. The weight setting can dynamically adapt to the saliency of different perspectives to ensure that the fused image is rich in details and has no obvious transition marks at the border. Through edge sharpening processing, the edge details of the image are enhanced, making the image clearer in visual performance, meeting the needs of virtual reality, face recognition, animation production, etc. for high-quality images. In addition, the computational complexity of the entire fusion process is low, which is suitable for real-time processing scenarios, further expanding the practical application scope of the method.

[0027] The value range of the pixel response degree value k to the correction force in S3 is 0.1 to 1.0. In the above embodiment, when optimizing the multi-view image, the method of the present invention uses the pixel response degree value k to the correction force to adjust the local pixel position. The value range of parameter k is set to 0.1 to 1.0; When k is close to 0.1, the correction is weak and suitable for areas with small errors or smooth changes in local features of the image. At this time, the adjustment of the pixel position is small to maintain the original details of the local area; when k is close to 1.0, the correction is strong and suitable for areas with large errors or significant features that need to be adjusted. At this time, the adjustment of the pixel position is large to ensure the significance of the optimization result. Through the dynamic adjustment of k, it can adapt to the optimization needs of different areas and ensure that the final generated multi-view image has higher consistency and realism.

[0028] The calculation formula of the intensity value F required to be applied when optimizing the target area in S3 is: ; Wherein, F represents the intensity value to be applied when optimizing the target area, G represents the grayscale value of the pixel point in the target area, p represents the horizontal position of the pixel point in the target area, and q represents the vertical position of the pixel point in the target area.

[0029] In the above embodiment, the present invention determines the intensity value F required for optimization by calculating the gray value gradient of the pixels in the target area. Formula The sum of the rates of change of the gray value G at the horizontal position p and the vertical position q is defined; Grayscale gradient calculation: Indicates the grayscale change rate of the target area pixels in the horizontal direction. Represents the grayscale change rate in the vertical direction. The sum of the two reflects the overall intensity of the grayscale change of the pixel. Optimization intensity determination: By calculating F, the intensity value that needs to be applied to the target area during the optimization process can be determined, thereby guiding the adjustment of the pixel position or grayscale value of the target area, so that the optimized image is smoother and more consistent. This method accurately reflects the change characteristics of the pixels in the target area through gradient calculation, so that the optimization process can focus on the significant change areas (such as edges or textures), thereby improving the image optimization effect. The present invention determines the optimization intensity by calculating the grayscale gradient, which can significantly improve the optimization effect on the target area. Compared with the traditional fixed intensity value optimization method, the gradient-based optimization method can dynamically adjust the optimization intensity according to different grayscale changes, thereby achieving a better balance between detail retention and image smoothing. In addition, the gradient calculation method is simple and efficient, and is suitable for real-time processing scenarios. The optimized image is clearer in the edge area, and the overall visual effect is more natural, which is suitable for multiple application scenarios such as multi-view generation, virtual reality and face reconstruction.

[0030] The facial feature points in S1 are extracted based on the key feature areas of the face, including the corners of the eyes, the tip of the nose, the corners of the mouth, and the chin area. In the above embodiment, the method of the present invention extracts facial feature points from the corners of the eyes, the tip of the nose, the corners of the mouth, and the chin area by analyzing the key feature areas of the input face image. Specifically, the following steps are included: Key area positioning: Through the face detection algorithm, the overall outline of the face is located and key areas (such as eyes, nose, mouth and chin areas) are accurately segmented.

[0031] Feature point extraction: Use specific algorithms (such as Haar features, Dlib key point detectors, or deep learning algorithms) to extract feature points in key areas to ensure the accuracy and robustness of key points. For example: Eye corners: locate the inner and outer corners of the eyes to determine the direction and position of the eyes; Nose tip: extract the center points of the nose bridge and nose tip as the reference for the geometric center of the face; Mouth corners: detect the left and right corners of the lips to describe the mouth features; Chin: extract the tip of the chin to help define the lower contour of the face.

[0032] Point correction and optimization: geometric correction is performed on the initially extracted feature points to ensure that the position of the points conforms to the actual facial structure. For example, symmetry constraints are used to ensure the alignment of the left and right feature points. By extracting feature points in the above key feature areas, high-precision geometric information can be provided for subsequent 3D model construction.

[0033] The value range of the angle θ rotated around a certain axis in S2 is [-180°, 180°], and image sequences of different target perspectives are generated by adjusting the angle. In the above embodiment, the present invention generates face images of multiple perspectives by adjusting the value range of the rotation angle θ. The specific working process is as follows: Initial position setting: Set the initial perspective of the 3D face model to the default position, usually the frontal perspective (θ=0°).

[0034] Angle adjustment mechanism: Set the range of the rotation angle θ to [-180°, 180°], covering the complete rotation space.

[0035] When θ>0°, it means that the model rotates clockwise; when θ<0°, it means that the model rotates counterclockwise.

[0036] View sequence generation: By uniformly or non-uniformly sampling θ within a specified range (for example, every 10° or dynamically adjusting the sampling interval according to specific needs), the rotation angle of the model is gradually adjusted to generate the corresponding target view image sequence.

[0037] Sequence image integration: Arrange the generated target perspective images in sequence for subsequent application scenarios (such as 3D display or video rendering). By setting the full angle range [-180°, 180°], the 3D face model can be displayed in all directions, from the front, side to the back, to generate a variety of perspective images.

[0038] The present invention can generate image sequences of different perspectives in an all-round range by clearly setting the value range of the rotation angle θ. Compared with the traditional technical solution that only generates limited perspectives, the present method can provide more comprehensive perspective coverage, and is particularly suitable for dynamic scene display in the fields of virtual reality, three-dimensional display and face recognition. In addition, the flexible angle adjustment mechanism makes the target perspective image generation more accurate, and can be customized according to different scene requirements (such as specific perspectives or perspective intervals), which improves the applicability and flexibility of the method.

[0039] In S4, the edge sharpening of the fused image is achieved through the Laplace operator. The specific formula is: ; Among them, L′ represents the brightness value of a pixel in the image after Laplace operator processing, L represents the brightness value of a pixel in the input image, and λ represents the sharpening coefficient. Represents the gradient operator.

[0040] In the above implementation, the present invention performs edge sharpening processing on the fused image by using the Laplacian operator to enhance the detail performance of the image. The specific principle is as follows: The role of Laplace operator: Laplace operator It is a second-order gradient operator used to capture areas in an image where brightness changes rapidly (such as edges). It can effectively highlight edge features in an image while having little impact on smooth areas.

[0041] Sharpening formula: Sharpening formula is the sharpening factor, which controls the strength of the sharpening effect: When λ>0, edge features are enhanced to make the image clearer; when λ is smaller, the sharpening effect is weaker; when λ is larger, the sharpening effect is stronger, but noise may be introduced.

[0042] Implementation process: By calculating the Laplacian operator value of each pixel in the image, multiplying it by the sharpening coefficient and adding it to the original pixel brightness value, the sharpened brightness value L' is obtained. This process is performed at the pixel level and the entire image is processed one by one, and finally a sharpened image with enhanced edge features is generated. This method can effectively improve the contrast of the image edge and make the visual effect clearer. It is especially suitable for optimizing the details of the fused multi-view images.

[0043] Using the Laplacian operator to sharpen the edges of the fused image can significantly improve the expressiveness of image details and edge clarity. Compared with traditional sharpening methods (such as the Sobel operator or high-pass filtering), the Laplacian operator has higher edge capture accuracy and is suitable for processing blurred areas in the fused multi-view images that may be caused by parallax optimization. By adjusting the sharpening coefficient λ, the sharpening intensity can be flexibly controlled to meet different application requirements. The sharpened image has a more impactful visual effect and meets the requirements for high-quality images in fields such as virtual reality, 3D reconstruction, and image recognition.

[0044] A system for generating a multi-view image of a human face based on a single view is also provided, which is used to implement the steps of the method for generating a multi-view image of a human face based on a single view, and the system comprises: A feature extraction module is used to extract feature points based on a single input face image and construct a 3D face model; A view transformation module connected to the feature extraction module is used to perform view transformation on the three-dimensional face model to generate face images from multiple perspectives; An image optimization module connected to the view transformation module is used to optimize the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax; The image integration module connected to the image optimization module is used to integrate the optimized multi-view images to generate a final multi-view image.

[0045] In the above implementation, the system realizes the complete process of generating multi-view face images based on a single view through modular design. The specific working principle is as follows: Feature extraction module: Use specific algorithms (such as Dlib, deep learning models, etc.) to analyze the input single face image and extract key feature points (including the corners of the eyes, the tip of the nose, the corners of the mouth, and the chin). Build a 3D face model based on the extracted feature points, and use a geometric fitting algorithm (such as the least squares method) to optimize the accuracy of the model to ensure the authenticity and stability of the 3D model.

[0046] View transformation module: Based on the constructed 3D face model, the rotation angle is adjusted around the specified rotation axis (such as the Z axis), and the rotation matrix is ​​used to generate images of different perspectives. By setting the angle range (such as [-180°, 180°]) and the sampling interval (such as every 10°), a sequence of face images with multiple perspectives is generated.

[0047] Image optimization module: Optimize the details of the generated multi-view images, including adjusting the inconsistent areas of the three-dimensional spatial structure caused by parallax. Use optimization algorithms (such as pixel adjustment formula Δd=k×F, gradient calculation formula ) Refine the target area to reduce inconsistencies.

[0048] Image integration module: weighted fusion of optimized multi-view images using the weighted average formula Calculate pixel values, remove splicing traces, and generate a smooth and consistent integrated image. Perform edge sharpening on the integrated image (e.g., using the Laplace operator). ) to enhance the expression of details and ultimately output high-quality multi-view face images.

[0049] Through modular design, this system can efficiently complete the whole process from single-view face image to multi-view image generation. Each module works independently, which is easy to upgrade or replace. For example, the feature extraction algorithm can be improved or the process can be optimized to adapt to different application scenarios. Through view transformation, optimization and integration processing, the multi-view face images generated by this system have higher visual consistency and realism, which is suitable for applications such as virtual reality, identity recognition and 3D animation production. It supports different algorithm combinations to meet diverse needs, such as using different optimization strategies or integration methods to enhance image expression.

[0050] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for generating multi-view images of a face based on a single view, characterized in that: The method comprises: S1, extracting feature points based on a single input face image and constructing a 3D face model; S2, performing view transformation on the three-dimensional face model to generate face images from multiple perspectives; S3, optimizing the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax, including taking the distorted area in the image as the optimization target, simulating local adjustment of the image, redistributing the pixel positions based on the significant feature points, and updating the optimized parallax-corrected image. The specific formula is: Δd=k×F; Wherein, Δd represents the pixel offset required to adjust the inconsistent three-dimensional spatial structure area caused by parallax, k represents the pixel response value to the correction strength, and F represents the strength value required to be applied when optimizing the target area; S4. Integrate the optimized multi-view images to generate a final multi-view image.

2. The method for generating multi-view images of a face based on a single view according to claim 1, characterized in that: S1 includes extracting facial feature points from the input image, fitting the feature points to generate a three-dimensional surface model, optimizing the parameters by the least squares method, and outputting a three-dimensional model generated based on the feature points. The specific formula is: ; Among them, z represents the depth value of the facial feature point, x and y represent the horizontal and vertical positions of the facial feature point in the two-dimensional plane image, a represents the influence of x on the depth z, b represents the influence of y on the depth z, and c represents the overall offset.

3. The method for generating multi-view images of a face based on a single view according to claim 1, characterized in that: The S2 includes: The initial position of the 3D face model is set as the default perspective, and the model is rotated around different axes to generate the target perspective. The specific formula is: ； Among them, u, v, w represent the horizontal coordinate, vertical coordinate, and depth coordinate of a point in the three-dimensional face model, respectively; u′, v′, w′ represent the new horizontal position, vertical position, and depth position of the point in the three-dimensional space after the rotation transformation, respectively; and θ represents the angle of rotation around a certain axis.

4. The method for generating multi-view images of a face based on a single view according to claim 1, characterized in that: The S4 includes: The weights are set for the multi-view images by region, and multiple images are fused by pixel to generate a full-view consistency image. The specific formula is: ; Where I represents the final value of a pixel of the target view image when integrating multi-view images, n represents the number of view images generated and involved in fusion, i represents the index number of the multi-view image, and h represents the index number of the multi-view image. i Indicates the weight of the i-th view in the fusion process, I i Represents the gray value of a pixel at a corresponding position in the i-th perspective image; Perform edge sharpening on the fused image.

5. The method for generating multi-view images of a face based on a single view according to claim 1, characterized in that: The value k of the pixel response degree to the correction force in S3 ranges from 0.1 to 1.

0.

6. The method for generating multi-view images of a face based on a single view according to claim 1, characterized in that: The calculation formula of the intensity value F that needs to be applied when optimizing the target area in S3 is: ; Wherein, F represents the intensity value to be applied when optimizing the target area, G represents the grayscale value of the pixel point in the target area, p represents the horizontal position of the pixel point in the target area, and q represents the vertical position of the pixel point in the target area.

7. The method for generating multi-view images of a face based on a single view according to claim 2, characterized in that: The facial feature point extraction in S1 is based on key feature areas of the face, including the corners of the eyes, the tip of the nose, the corners of the mouth and the chin area.

8. The method for generating multi-view images of a face based on a single view according to claim 3, characterized in that: The value range of the angle θ of rotation around a certain axis in S2 is [-180°, 180°], and image sequences with different target viewing angles are generated by adjusting the angle.

9. The method for generating multi-view images of a face based on a single view according to claim 4, characterized in that: In S4, edge sharpening of the fused image is performed by Laplace operator, and the specific formula is: ; Among them, L′ represents the brightness value of a pixel in the image after Laplace operator processing, L represents the brightness value of a pixel in the input image, λ represents the sharpening coefficient, Represents the gradient operator.

10. A system for generating multi-view images of a face based on a single view, used to implement the steps of the method for generating multi-view images of a face based on a single view according to any one of claims 1 to 9, characterized in that: The system comprises: A feature extraction module is used to extract feature points based on a single input face image and construct a 3D face model; A view transformation module connected to the feature extraction module is used to perform view transformation on the three-dimensional face model to generate face images from multiple perspectives; An image optimization module connected to the view transformation module is used to optimize the generated multi-view images to reduce the inconsistency of the three-dimensional spatial structure caused by parallax; The image integration module connected to the image optimization module is used to integrate the optimized multi-view images to generate a final multi-view image.

Images