A facial profile correction result prediction system and apparatus

Abstract

This invention discloses a facial profile correction result prediction system and device, belonging to the field of medical image processing and deep learning technology. The system includes: a dataset construction module for correcting an initial facial profile image using a correction algorithm to obtain a target facial profile image and constructing a paired dataset; a dataset preprocessing module for randomly cropping each pair of paired data to obtain a standard paired training dataset; a model construction module for constructing a generator and a discriminator, and constructing an adversarial network model to be trained based on the generator and discriminator; a model training module for inputting the standard paired training dataset into the adversarial network model to be trained, and iteratively training it using a training strategy to obtain a target adversarial network model; and a result prediction module for inputting the facial profile image to be predicted into the target adversarial network model and outputting the facial profile correction prediction result. Therefore, this invention can improve the accuracy of postoperative facial profile correction prediction results.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and deep learning technology, and in particular to a facial profile correction result prediction system and device. Background Technology

[0002] Dentofacial deformities are common maxillofacial developmental disorders that can cause malocclusion and alter facial contours. Orthognathic surgery is currently the main treatment for dentofacial deformities. However, there is still considerable uncertainty in predicting the postoperative changes in facial soft tissue appearance before orthognathic surgery. How to accurately and reliably predict the results of facial profile correction has become a technical problem that needs further research.

[0003] Currently, existing technologies for predicting postoperative facial appearance mainly employ two-dimensional image processing. Operators use general image processing software to perform local deformation stretching and pixel-level manual editing on the maxillofacial region of the patient's preoperative profile image to generate a predicted image for facial profile correction. However, existing technologies do not establish an objective quantitative mapping law between changes in bony structures and the appearance response of soft tissues during the deformation and editing process of the maxillofacial region. Furthermore, they rely on the operator's subjective aesthetic sense, making it difficult for the prediction method of facial profile correction results to reflect the actual anatomical structure changes, resulting in poor accuracy of the postoperative effect diagram. Summary of the Invention

[0004] This invention proposes a facial profile correction result prediction system and device, which solves the problem that existing technologies using two-dimensional image processing methods to edit preoperative images at the pixel level to generate postoperative effect diagrams often fail to reflect the true anatomical structure changes, resulting in poor accuracy of the postoperative effect diagrams. This invention improves the accuracy of postoperative facial profile correction prediction results by learning the mapping relationship between facial structure and appearance, reducing the influence of random factors on the generated postoperative effect diagrams.

[0005] To achieve the above objectives, this invention provides a facial profile correction result prediction system, including a dataset construction module, a dataset preprocessing module, a model construction module, a model training module, and a result prediction module; wherein: the dataset construction module is used to perform image correction on a pre-acquired initial facial profile image using a preset correction algorithm to obtain a target facial profile image, and construct a paired dataset based on the mapping relationship between the initial facial profile image and the target facial profile image; the dataset preprocessing module is used to obtain a standard paired training dataset by randomly cropping each set of paired data in the paired dataset; The model building module is used to construct a generator by cascading a global generator network and a local enhancement network, construct a discriminator based on several sets of parameter-independent sub-discriminator networks, and construct a trainable adversarial network model based on the generator and the discriminator. The model training module is used to input the standard paired training dataset into the trainable adversarial network model, iteratively update the model parameters of the trainable adversarial network model through a preset training strategy until the model parameters meet the preset training requirements, and obtain the target adversarial network model. The result prediction module is used to input the facial profile image to be predicted into the target adversarial network model, and output the facial profile correction prediction result through the target adversarial network model.

[0006] This invention proposes a facial profile correction result prediction system. A dataset construction module builds paired datasets based on initial facial profile images and corresponding correction target images, providing a data foundation with a realistic mapping relationship for subsequent model training. Then, a dataset preprocessing module performs standardization processing such as random cropping on the paired data, unifying data input specifications and reducing interference from data collection differences on model training. Next, a model construction module uses a generator with a global generative network cascaded with a local augmentation network, combined with a discriminator composed of multiple independent sub-discriminator networks to build an adversarial network, balancing global contour rationality and local detail restoration capabilities at the network structure level. Furthermore, a model training module iteratively optimizes network parameters through a preset training strategy, enabling the model to learn the realistic mapping rules of facial profile correction. Finally, a result prediction module outputs the correction prediction result through the trained model. Thus, based on the synergistic effect of each module, an intrinsic mapping relationship between facial structure changes and appearance correction is established in a data-driven manner. A deep learning model learns the corresponding mapping relationship and generates facial profile correction results that conform to objective correction rules, reducing the influence of random factors on the generated post-operative effect diagram and improving the accuracy of post-operative facial profile correction prediction results.

[0007] Furthermore, the dataset construction module includes a validity screening unit, an image correction unit, and a data pairing unit; wherein: the validity screening unit is used to collect initial facial profile images, and based on preset validity screening indicators, remove initial facial profile images that are occluded, blurred, or distorted to obtain valid initial facial profile images; the image correction unit is used to perform contour correction and texture restoration on deformed areas in the valid initial facial profile images based on preset correction algorithms to obtain target facial profile images that are mapped one-to-one with the valid initial facial profile images; the data pairing unit is used to pair and bind each pair of corresponding valid initial facial profile images and target facial profile images to construct a paired dataset.

[0008] In the above approach, initial facial profile images undergo validity screening to remove unqualified images. Then, professional contour correction is performed, and the corresponding images are paired and bound to form a paired dataset. Thus, validity screening eliminates interference from low-quality images on model training, ensuring the basic quality of the training data. Simultaneously, the professionally corrected images form a one-to-one mapping relationship, enabling the model to learn correction rules that conform to objective standards. This avoids training bias caused by data with no corresponding relationship, ultimately constructing a high-quality paired dataset with accurate mapping relationships. This provides a reliable data foundation for subsequent model learning and training, ensuring the model can learn reliable pre- and post-operative mapping relationships, and helping to improve the accuracy of post-operative facial profile correction prediction results.

[0009] Furthermore, the dataset preprocessing module includes an image cropping unit, an image preprocessing unit, and a data integration unit; wherein: the image cropping unit is used to randomly crop the effective initial facial profile images and target facial profile images in each set of initial pairing data according to a preset ratio within a preset cropping area to obtain several local images; the image preprocessing unit is used to restore the size of the several local images to be consistent with the size of the target facial profile image through a preset magnification algorithm, and to perform resolution standardization and pixel value normalization processing on the several local images after size restoration to obtain several sets of target pairing data; the data integration unit is used to integrate each set of initial pairing data and several sets of target pairing data into a standard pairing training dataset.

[0010] In the above scheme, synchronous random cropping is performed on each pairing data set, and after size adjustment, resolution standardization, and pixel value normalization, a standard paired training dataset is output. Thus, by performing synchronous processing on the paired images, the spatial correspondence between the input and target images is ensured to remain unchanged, avoiding pairing misalignment from affecting model training performance. Simultaneously, standardizing image specifications and pixel distribution reduces data differences caused by different acquisition conditions and minimizes interference from irrelevant factors in the model learning process. Ultimately, uniform and accurately paired training data is obtained, providing a reliable data foundation for subsequent model training. This ensures the model can focus on learning the correction mapping rules of facial profiles, improving the accuracy of postoperative facial profile correction prediction results.

[0011] Furthermore, the model construction module includes a generator construction unit, a discriminator construction unit, and a trainable adversarial network model construction unit. The generator construction unit is used to construct a global generator network based on a preset first end-to-end convolutional architecture and a preset first resolution layer, and to construct a local enhancement network based on a preset second end-to-end convolutional architecture and a preset second resolution layer. The global generator network and the local enhancement network are cascaded to obtain a generator. The resolution of the preset first resolution layer is smaller than the resolution of the preset second resolution layer. The discriminator construction unit is used to set independent and different network configuration parameters for each of several initial sub-discriminator networks with the same network structure to obtain several sets of sub-discriminator networks, and to combine the several sets of sub-discriminator networks into a discriminator. The trainable adversarial network model construction unit is used to construct a target loss function based on the generator and the discriminator, and to construct a trainable adversarial network model based on the target loss function.

[0012] In the above scheme, a generator is constructed by cascading a global generation network and a local enhancement network. The global generation network receives the input image and generates global structural information at a first resolution level, while the local enhancement network receives this global structural information and generates the target image at a higher second resolution level. This coarse-to-fine hierarchical generation strategy allows the generator to first grasp the overall facial contour and spatial layout, and then gradually supplement high-frequency details. Several sub-discriminator networks are then constructed, each with the same network structure and independent parameters. Each sub-discriminator network receives image data at different resolution levels and performs discrimination, allowing the discriminator to supervise the generation results at multiple scales. Finally, the generator, discriminator, and the constructed target loss function are combined to construct the adversarial network model to be trained. Thus, during training, the model continuously optimizes the generation quality through the interplay between the generator and the discriminator. The generator refines the image step by step through the cascaded structure, while the discriminator focuses on both the overall picture and the local details through multi-scale supervision. The collaboration between the two forces the generated image to possess both a reasonable global structure and realistic local details, thereby improving the accuracy of postoperative facial profile correction prediction results.

[0013] Furthermore, the generator construction unit includes: the global generation network is used to downsample the input image step by step according to a preset first end-to-end convolutional architecture to obtain an initial global semantic feature map, perform residual propagation on the initial global semantic feature map based on a preset nonlinear activation function to obtain a target global semantic feature map, and perform step by step upsampling on the target global semantic feature map based on a preset first resolution level to obtain an initial facial profile correction prediction map.

[0014] In the above scheme, the global generation network adopts a preset first end-to-end convolutional architecture, sequentially performing stepwise downsampling, residual feature propagation, and stepwise upsampling operations. First, the initial global semantic feature map of the input image is extracted through downsampling, then the target global semantic feature map is obtained through nonlinear activation and residual propagation, and finally, the initial facial profile correction prediction map is generated at a low-resolution level through upsampling. Thus, by progressively compressing the feature map size through stepwise downsampling, the global semantics and overall contour information of the image are effectively extracted. The residual propagation mechanism can enhance the feature expression ability while maintaining the stability of the feature dimension, and at the same time alleviate the gradient degradation problem of deep networks. The upsampling reconstruction at the low-resolution level can grasp the rationality of the overall facial structure, and finally output an initial facial profile correction prediction map with accurate global contour layout, providing a reliable global structural prior for subsequent high-resolution detail generation, effectively improving the accuracy of postoperative facial profile correction prediction results.

[0015] Furthermore, the generator construction unit includes: the local enhancement network is used to downsample the input image step by step according to a preset second end-to-end convolutional architecture, extract local feature maps, fuse the local feature maps and the initial facial profile correction prediction map to obtain an initial fused feature map, perform convolutional transformation on the initial fused feature map to obtain a target fused feature map, and perform step-by-step upsampling on the target fused feature map based on a preset second resolution level to obtain a target facial profile correction prediction map.

[0016] In the above scheme, the local enhancement network adopts a pre-defined second end-to-end convolutional architecture. First, it extracts local feature maps from the input image through progressive downsampling. Then, it fuses these local feature maps with the initial facial profile correction prediction map output by the global generator network to obtain an initial fused feature map. After convolutional transformation to obtain the target fused feature map, it progressively upsamples at high-resolution levels to obtain the target facial profile correction prediction map. Thus, by extracting local texture and edge detail features from the image through downsampling and fusing them with global structural features, a coordinated expression of global layout and local details can be achieved. The high-resolution convolutional transformation and upsampling further refine high-frequency detail information, compensating for the insufficient detail expression of a single generator network at high resolution. Ultimately, it can generate a high-resolution correction prediction image that combines overall structural consistency and local texture realism, helping to improve the accuracy of postoperative facial profile correction prediction results.

[0017] Furthermore, the discriminator construction unit includes: the sub-discriminator network for acquiring the resolution of the input image; if the resolution is less than or equal to a preset first resolution threshold, then the overall structural information of the input image is discriminated to generate a first discrimination result; if the resolution is greater than the preset first resolution threshold and less than a preset second resolution threshold, then the mesoscale structural features of the input image are discriminated to generate a second discrimination result; if the resolution is greater than or equal to the preset second resolution threshold, then the local texture and edge contours of the input image are discriminated to generate a third discrimination result.

[0018] In the above scheme, multiple sets of parameter-independent sub-discriminator networks are constructed. Images of different resolutions are input into the corresponding sub-discriminator networks through multi-scale downsampling, and the corresponding first, second, or third discrimination results are output. Thus, by utilizing the different receptive field ranges of the sub-discriminator networks at different resolutions, the authenticity of the generated images is verified from multiple levels, including global structure, mesoscale contours, and local details. This ensures that the images generated by the subsequent model conform to the features of real images in both overall structure and local details, helping to improve the accuracy of postoperative facial profile correction prediction results.

[0019] Furthermore, the model training module includes a forward propagation unit, an optimization target construction unit, and a model parameter update unit; wherein: the forward propagation unit is used to obtain a training dataset from the standard paired training dataset, input the training dataset into the adversarial network model to be trained, and obtain the first discrimination result, the second discrimination result, or the third discrimination result; the optimization target construction unit is used to calculate and generate adversarial loss and feature matching loss based on the target loss function, according to the first discrimination result, the second discrimination result, or the third discrimination result, and perform a weighted combination of the adversarial loss and the feature matching loss to obtain the optimization target; the model parameter update unit is used to alternately update the generator and discriminator in the adversarial network model to be trained according to the optimization target until the model parameters of the adversarial network model to be trained meet the preset training requirements, thereby obtaining the target adversarial network model.

[0020] In the above scheme, training data is first input into the model to be trained to obtain the discrimination results at each level. Then, adversarial loss and feature matching loss are calculated based on the target loss function. After weighted combination, the overall optimization target is obtained. Finally, the parameters of the generator and discriminator are alternately updated according to the optimization target until the model parameters meet the preset training requirements, thus obtaining the target adversarial network model. Therefore, multi-scale discrimination information is obtained through forward propagation, providing multi-dimensional basis for loss calculation. The composite loss function can simultaneously constrain the visual realism and feature-level consistency of the generated image. The alternating update training method enables adversarial game and collaborative optimization between the generator and discriminator, ultimately enabling the model to converge stably and learn accurate facial profile correction mapping relationships. This constructs a reliable target adversarial network model, thereby improving the accuracy of postoperative facial profile correction prediction results.

[0021] Furthermore, the result prediction module includes a to-be-predicted image acquisition unit and a prediction result output unit; wherein: the to-be-predicted image acquisition unit is used to acquire a to-be-predicted facial profile image; the prediction result output unit is used to input the to-be-predicted facial profile image into the target adversarial network model, and perform feature extraction and multi-scale feature fusion on the to-be-predicted facial profile image through the generator in the target adversarial network model to generate a corresponding facial profile correction prediction image, and output the facial profile correction prediction image as the facial profile correction prediction result.

[0022] In the above scheme, a facial profile image to be predicted is first acquired. Then, the generator in the target adversarial network model extracts features and fuses multi-scale features from the facial profile image. The trained generator automatically generates a facial profile correction prediction image, which is then output as the correction prediction result. Thus, by automatically extracting multi-dimensional features from the image to be predicted using the target model, and automatically generating the corresponding correction result based on the mapping relationship between preoperative facial structure features and postoperative appearance changes learned during network training, pixel-level editing is eliminated, ensuring that the prediction result follows the objective correction rules learned by the model. This helps improve the accuracy of postoperative facial profile correction prediction results.

[0023] This invention also provides a facial profile correction result prediction device, the device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor applies the facial profile correction result prediction system when executing the computer program. Attached Figure Description

[0024] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the module structure of a facial profile correction result prediction system provided in a certain embodiment of the present invention; Figure 2 This is a schematic diagram of a single data pair of a facial profile correction result prediction system proposed in a certain embodiment of the present invention. Figure 3 This is a schematic diagram of random cropping for a facial profile correction result prediction system proposed in a certain embodiment of the present invention. Figure 4 This is a schematic diagram of the generator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention. Figure 5 This is a schematic diagram of the discriminator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; Figure 6 This is a schematic diagram of the prediction simulation process of a facial profile correction result prediction system according to a certain embodiment of the present invention; Figure 7 This is a schematic diagram of the prediction result output of a facial profile correction result prediction system proposed in a certain embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0028] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0029] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0030] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0031] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0032] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0033] To address the problem that existing technologies, which use two-dimensional image processing to edit preoperative images at the pixel level to generate postoperative effect diagrams, often fail to reflect the true changes in anatomical structures, resulting in inaccurate postoperative effect diagrams, see [reference needed]. Figure 1 , Figure 1 This is a schematic diagram of the module structure of a facial profile correction result prediction system provided in a certain embodiment of the present invention. Figure 1 As shown in the figure, this invention proposes a facial profile correction result prediction system, including a dataset construction module 101, a dataset preprocessing module 102, a model construction module 103, a model training module 104, and a result prediction module 105; wherein: the dataset construction module 101 is used to perform image correction on the pre-collected initial facial profile image using a preset correction algorithm to obtain a target facial profile image, and construct a paired dataset based on the mapping relationship between the initial facial profile image and the target facial profile image; the dataset preprocessing module 102 is used to obtain a standard paired training dataset by randomly cropping each set of paired data in the paired dataset. The model building module 103 is used to cascade a global generator network and a local enhancement network to build a generator, build a discriminator based on several sets of parameter-independent sub-discriminators, and build a training adversarial network model based on the generator and the discriminator. The model training module 104 is used to input the standard paired training dataset into the training adversarial network model, iteratively update the model parameters of the training adversarial network model through a preset training strategy until the model parameters meet the preset training requirements, and obtain the target adversarial network model. The result prediction module 105 is used to input the facial profile image to be predicted into the target adversarial network model, and output the facial profile correction prediction result through the target adversarial network model.

[0034] One possible implementation involves using high-resolution SLR cameras and other professional imaging equipment to acquire original facial profile images of patients with various skeletal maxillofacial deformities. The acquired original images are then screened for validity, including image integrity, resolution, illumination uniformity, shooting angle consistency, and clarity. Low-quality images with image occlusion, blurring, illumination distortion, and shooting angle deviation are removed, leaving valid initial facial profile images that meet unified imaging standards. Then, based on existing standardized correction algorithms adapted to the orthognathic surgery maxillofacial deformity correction logic, contour reshaping and simultaneous soft tissue texture repair are performed on deformed areas such as mandibular retrusion and underbite within the valid images. This generates a unique target facial profile image corresponding to each valid initial image. The initial images with one-to-one mapping relationships are bound and archived with the corrected target images to form a paired dataset with clinical anatomical correspondence rules. Then, all paired image samples within the paired dataset are retrieved, and synchronous random cropping is performed on both the initial facial profile image and the target facial profile image within the same group. After cropping different areas of the image, the cropped local images are restored to a uniform standard size using an interpolation scaling algorithm. Simultaneously, existing preprocessing methods are used to perform a series of standardization operations on all images, such as size unification, resolution standardization, and pixel value normalization. The newly added expanded samples generated by synchronous transformation are integrated and summarized with the original paired samples to form a standard paired training dataset with uniform specifications and sufficient sample quantity. Furthermore, an improved pix2pixHD conditional generative adversarial network is used to construct the main network structure. Then, a two-level sequentially cascaded generator structure is built. The low-resolution global generation network relies on multiple convolutional layers to downsample and extract global semantic features from the input image. Stable global features are passed through multiple residual blocks, and then progressively upsampled to output a low-resolution initial facial profile correction prediction map containing the complete overall facial contour. The high-resolution local enhancement network independently extracts local feature maps from the input image. These local feature maps are then compared with the low-resolution initial facial profile correction map output by the global generation network. The positive prediction images are fused, and the fused features are refined by residual convolution to refine the jaw and facial edges and soft tissue textures. Then, a high-definition and complete target facial profile correction prediction image is output through progressive upsampling. At the same time, a multi-scale discriminator is built, and multiple sets of sub-discriminator networks with consistent network structures and independent parameters are prepared. Each sub-discriminator network is adapted to the real and fake image discrimination of images at different resolution levels. The hierarchical cascade generator and multiple sets of independent sub-discriminator networks are combined into a complete network body. Then, combined with the target loss function adapted to the facial profile image generation task, a complete adversarial network model to be trained is built. After obtaining the adversarial network model to be trained, the model is loaded according to the existing deep learning development framework. An adversarial training strategy of alternating parameter updates by the generator and discriminator is adopted to carry out iterative optimization. During the training process, the multi-scale sub-discriminator network outputs discrimination results at different levels. The adversarial loss and feature matching loss are calculated synchronously according to the preset target loss function. The two types of losses are weighted and fused to obtain the overall optimization target. All parameters inside the network are continuously updated with the help of the gradient backpropagation mechanism. The network weights at each stage are stored at regular intervals during the training cycle, and the image generation effect of the model on the validation subset is continuously monitored until the structural rationality and detail reproduction of the model output image reach the preset training requirements. The iteration is terminated, and the network weights with the best validation effect are selected and solidified to obtain the target adversarial network model. Finally, a brand-new preoperative facial profile image to be predicted is received in real time. The image is first subjected to standardized preprocessing operations that are completely consistent with those performed during the model training phase, but the random cropping and augmentation operations performed during the training phase are not performed. The processed facial profile image to be predicted is then fed into the trained target adversarial network model to perform forward inference operations. The target adversarial network model autonomously extracts the global maxillofacial skeletal contour features and local soft tissue texture features contained in the facial profile image to be predicted. Combined with the intrinsic mapping rules of preoperative and postoperative profiles learned during the training phase, it automatically generates a high-resolution facial profile correction prediction image as the facial profile correction prediction result output.

[0035] This invention proposes a facial profile correction result prediction system. A dataset construction module builds paired datasets based on initial facial profile images and corresponding correction target images, providing a data foundation with a realistic mapping relationship for subsequent model training. Then, a dataset preprocessing module performs standardization processing such as random cropping on the paired data, unifying data input specifications and reducing interference from data collection differences on model training. Next, a model construction module uses a generator with a global generative network cascaded with a local augmentation network, combined with a discriminator composed of multiple independent sub-discriminator networks to build an adversarial network, balancing global contour rationality and local detail restoration capabilities at the network structure level. Furthermore, a model training module iteratively optimizes network parameters through a preset training strategy, enabling the model to learn the realistic mapping rules of facial profile correction. Finally, a result prediction module outputs the correction prediction result through the trained model. Thus, based on the synergistic effect of each module, an intrinsic mapping relationship between facial structure changes and appearance correction is established in a data-driven manner. A deep learning model learns the corresponding mapping relationship and generates facial profile correction results that conform to objective correction rules, reducing the influence of random factors on the generated post-operative effect diagram and improving the accuracy of post-operative facial profile correction prediction results.

[0036] In a preferred embodiment, the dataset construction module 101 includes a validity screening unit 201, an image correction unit 202, and a data pairing unit 203; wherein: the validity screening unit 201 is used to acquire initial facial profile images, and based on preset validity screening indicators, remove initial facial profile images that are occluded, blurred, or distorted to obtain valid initial facial profile images; the image correction unit 202 is used to perform contour correction and texture restoration on deformed areas in the valid initial facial profile images based on a preset correction algorithm to obtain target facial profile images that are mapped one-to-one with the valid initial facial profile images; the data pairing unit 203 is used to pair and bind each pair of corresponding valid initial facial profile images and target facial profile images to construct a paired dataset.

[0037] For example, initial facial profile images are acquired by connecting to professional high-resolution SLR cameras or other image acquisition equipment. These initial facial profile images include preoperative facial images of different types of skeletal maxillofacial deformities. Based on validity screening criteria, each image is individually verified for completeness, resolution compliance, illumination uniformity, shooting angle consistency, and image clarity. In this embodiment, preset validity screening indicators represent the validity screening criteria. Unqualified images with maxillofacial region occlusion, blurred images, or color distortion are discarded, retaining only valid initial facial profile images. Then, see... Figure 2 , Figure 2 This is a schematic diagram of a single data pair of a facial profile correction result prediction system proposed in a certain embodiment of the present invention, as shown below. Figure 2 As shown, existing correction algorithms are used, with standardized correction logic formed based on orthognathic surgery specialty treatment guidelines and clinical experience. For each valid initial facial profile image, key soft tissue areas such as the mandibular border, chin, and facial contour are accurately located. Coordinated contour correction is performed on the bony contour abnormalities corresponding to deformities such as mandibular retrusion and underbite. Simultaneously, the soft tissue texture around the correction area is naturally repaired, thereby generating target facial profile images that correspond one-to-one with the valid initial facial profile images. Then, a unified sample identifier is configured for each set of corresponding valid initial facial profile images and target facial profile images, and a one-to-one mapping binding relationship is established between the two. The data is archived and stored according to a unified data format, resulting in several sets such as... Figure 2 The pairing data shown, Figure 2 The left side of the image represents the valid initial facial profile image, and the right side represents the corresponding target facial profile image, thus obtaining the paired dataset.

[0038] In the above approach, initial facial profile images undergo validity screening to remove unqualified images. Then, professional contour correction is performed, and the corresponding images are paired and bound to form a paired dataset. Thus, validity screening eliminates interference from low-quality images on model training, ensuring the basic quality of the training data. Simultaneously, the professionally corrected images form a one-to-one mapping relationship, enabling the model to learn correction rules that conform to objective standards. This avoids training bias caused by data with no corresponding relationship, ultimately constructing a high-quality paired dataset with accurate mapping relationships. This provides a reliable data foundation for subsequent model learning and training, ensuring the model can learn reliable pre- and post-operative mapping relationships, and helping to improve the accuracy of post-operative facial profile correction prediction results.

[0039] In a preferred embodiment, the dataset preprocessing module 102 includes an image cropping unit 301, an image preprocessing unit 302, and a data integration unit 303; wherein: the image cropping unit 301 is used to randomly crop the effective initial facial profile images and target facial profile images in each set of initial pairing data according to a preset ratio within a preset cropping area to obtain several local images; the image preprocessing unit 302 is used to restore the size of the several local images to be consistent with the size of the target facial profile image through a preset magnification algorithm, and to perform resolution standardization and pixel value normalization processing on the several local images after size restoration to obtain several sets of target pairing data; the data integration unit 303 is used to integrate each set of initial pairing data and several sets of target pairing data into a standard pairing training dataset.

[0040] For example, see Figure 3 , Figure 3 This is a schematic diagram of random cropping in a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown. Figure 3 As shown, initial paired data from the paired dataset is retrieved, and a synchronous random cropping operation is performed on the valid initial facial profile image and the corresponding target facial profile image within each group. The cropping process randomly selects cropping positions within a preset region containing the core facial contour features, based on a preset ratio range. In this embodiment, such as... Figure 3Local regions covering the mandibular border and key structures of the chin, such as the upper right or lower right region of the mid-profile image, are cropped. The cropping position and cropping ratio of two images within the same group are kept completely consistent, resulting in multiple sets of spatially corresponding local images. These cropped local images are then received, and existing interpolation and magnification algorithms are used to restore the size of all local images to the standard input size consistent with the original target facial profile image. Subsequently, resolution standardization is uniformly performed on all the restored images, adjusting them to a unified resolution parameter. Then, pixel values ​​are normalized and mapped to a unified numerical range, resulting in multiple sets of target pairing data with uniform specifications and accurate mapping relationships. Finally, the original initial pairing data and the multiple sets of target pairing data generated by cropping and augmentation are summarized and integrated, and the dataset is archived according to a unified data storage format and annotation rules, forming a standard pairing training dataset with consistent specifications and sufficient samples.

[0041] In the above scheme, synchronous random cropping is performed on each pairing data set, and after size adjustment, resolution standardization, and pixel value normalization, a standard paired training dataset is output. Thus, by performing synchronous processing on the paired images, the spatial correspondence between the input and target images is ensured to remain unchanged, avoiding pairing misalignment from affecting model training performance. Simultaneously, standardizing image specifications and pixel distribution reduces data differences caused by different acquisition conditions and minimizes interference from irrelevant factors in the model learning process. Ultimately, uniform and accurately paired training data is obtained, providing a reliable data foundation for subsequent model training. This ensures the model can focus on learning the correction mapping rules of facial profiles, improving the accuracy of postoperative facial profile correction prediction results.

[0042] In a preferred embodiment, the model construction module 103 includes a generator construction unit 401, a discriminator construction unit 402, and a trainable adversarial network model construction unit 403. The generator construction unit 401 is used to construct a global generator network based on a preset first end-to-end convolutional architecture and a preset first resolution layer, and to construct a local enhancement network based on a preset second end-to-end convolutional architecture and a preset second resolution layer. The global generator network and the local enhancement network are cascaded to obtain a generator. The resolution of the preset first resolution layer is less than the resolution of the preset second resolution layer. The discriminator construction unit 402 is used to set independent and different network configuration parameters for each of several initial sub-discriminator networks with the same network structure, thereby obtaining several sets of sub-discriminator networks, and then combining these sets of sub-discriminator networks into a discriminator. The trainable adversarial network model construction unit 403 is used to construct a target loss function based on the generator and the discriminator, and to construct a trainable adversarial network model based on the target loss function.

[0043] For example, see Figure 4 , Figure 4 This is a schematic diagram of the generator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown below. Figure 4 As shown, a coarse-to-fine hierarchical cascaded construction approach is adopted. First, a global generator network G1 is built based on a preset first end-to-end convolutional architecture and a lower first resolution layer. In this embodiment, the preset first end-to-end convolutional architecture adopts an encoder-residual bottleneck-decoder end-to-end convolutional structure. Then, a local enhancement network G2 is built based on a preset second end-to-end convolutional architecture and a higher second resolution layer. In this embodiment, the preset second end-to-end convolutional architecture adopts an encoding-residual transform-decoding convolutional structure. The global generator network G1 and the local enhancement network G2 are cascaded in the processing order to obtain the generator. See Figure 5 , Figure 5 This is a schematic diagram of the discriminator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown. Figure 5 As shown, multiple initial sub-discriminator networks are then constructed simultaneously. Each sub-discriminator network adopts the same basic network structure, and each sub-discriminator network consists of multiple layers of convolutional units, normalization units, and nonlinear activation units. The network output simultaneously generates true / false discrimination results and intermediate layer feature maps. Then, independent and non-shared network parameters are configured for each sub-discriminator network. Since each sub-discriminator network receives input images downsampled at different ratios, its effective receptive field on the original image differs, thus adapting to image discrimination tasks at different resolution levels. Finally, multiple sets of parameter-independent sub-discriminator networks are combined. These are combined to form a multi-scale discriminator, in which, , where n is the total number of sub-discriminators; Finally, combining the architectural characteristics of the generator and the multi-scale discriminator, a target loss function adapted to the facial profile correction task is constructed. In this embodiment, the target loss function includes two parts: Generative Adversarial Loss (GAN Loss) and Feature Matching Loss (Feature Matching Loss), which are used to constrain the visual realism of the generated image and the distribution consistency at the feature level, respectively. The generator, discriminator, and target loss function are integrated to build a complete adversarial network model to be trained.

[0044] In the above scheme, a generator is constructed by cascading a global generation network and a local enhancement network. The global generation network receives the input image and generates global structural information at a first resolution level, while the local enhancement network receives this global structural information and generates the target image at a higher second resolution level. This coarse-to-fine hierarchical generation strategy allows the generator to first grasp the overall facial contour and spatial layout, and then gradually supplement high-frequency details. Several sub-discriminator networks are then constructed, each with the same network structure and independent parameters. Each sub-discriminator network receives image data at different resolution levels and performs discrimination, allowing the discriminator to supervise the generation results at multiple scales. Finally, the generator, discriminator, and the constructed target loss function are combined to construct the adversarial network model to be trained. Thus, during training, the model continuously optimizes the generation quality through the interplay between the generator and the discriminator. The generator refines the image step by step through the cascaded structure, while the discriminator focuses on both the overall picture and the local details through multi-scale supervision. The collaboration between the two forces the generated image to possess both a reasonable global structure and realistic local details, thereby improving the accuracy of postoperative facial profile correction prediction results.

[0045] In a preferred embodiment, the generator construction unit 401 includes: a global generation network used to downsample the input image step by step according to a preset first end-to-end convolutional architecture to obtain an initial global semantic feature map, to perform residual propagation on the initial global semantic feature map based on a preset nonlinear activation function to obtain a target global semantic feature map, and to upsample the target global semantic feature map step by step based on a preset first resolution level to obtain an initial facial profile correction prediction map.

[0046] For example, see Figure 4 , Figure 4 This is a schematic diagram of the generator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown below. Figure 4 As shown, a coarse-to-fine hierarchical cascade construction approach is adopted. First, a global generation network G1 is built based on a preset first end-to-end convolutional architecture and a lower preset first resolution layer. In this embodiment, the preset first end-to-end convolutional architecture adopts an encoder-residual bottleneck-decoder end-to-end convolutional structure. In the encoder encoding stage, the input image is downsampled step by step through multiple convolutions with a stride of 2 to extract global semantic features of the image and compress the feature map space size. In the residual bottleneck stage, it is composed of multiple residual blocks stacked together. Each residual block contains two 3×3 convolutional layers. Feature residuals are propagated by identity skip connections, which stabilizes the feature map size, enhances feature expression ability, and alleviates the gradient degradation problem of deep networks. In the decoder decoding stage, the feature map resolution is restored step by step through multiple upsampling convolutions, and the initial facial profile correction prediction map of the low resolution layer is output.

[0047] In the above scheme, the global generation network adopts a preset first end-to-end convolutional architecture, sequentially performing stepwise downsampling, residual feature propagation, and stepwise upsampling operations. First, the initial global semantic feature map of the input image is extracted through downsampling, then the target global semantic feature map is obtained through nonlinear activation and residual propagation, and finally, the initial facial profile correction prediction map is generated at a low-resolution level through upsampling. Thus, by progressively compressing the feature map size through stepwise downsampling, the global semantics and overall contour information of the image are effectively extracted. The residual propagation mechanism can enhance the feature expression ability while maintaining the stability of the feature dimension, and at the same time alleviate the gradient degradation problem of deep networks. The upsampling reconstruction at the low-resolution level can grasp the rationality of the overall facial structure, and finally output an initial facial profile correction prediction map with accurate global contour layout, providing a reliable global structural prior for subsequent high-resolution detail generation, effectively improving the accuracy of postoperative facial profile correction prediction results.

[0048] In a preferred embodiment, the generator construction unit 401 includes: a local enhancement network for progressively downsampling the input image according to a preset second end-to-end convolutional architecture to extract local feature maps, fusing the local feature maps and the initial facial profile correction prediction map to obtain an initial fused feature map, performing convolutional transformation on the initial fused feature map to obtain a target fused feature map, and progressively upsampling the target fused feature map based on a preset second resolution level to obtain a target facial profile correction prediction map.

[0049] For example, see Figure 4 , Figure 4 This is a schematic diagram of the generator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown below. Figure 4 As shown, a local enhancement network G2 is built based on a preset second end-to-end convolutional architecture and a higher preset second resolution layer. In this embodiment, the preset second end-to-end convolutional architecture adopts an encoding-residual transformation-decoding convolutional structure. In the encoder encoding stage, local detail features of the input image are extracted, and the local features are fused with the initial facial profile correction prediction map output by the global generation network. In the residual transformation stage, the fused result is further processed by multiple residual convolutional transformations to learn the high-frequency texture and edge contour detail mapping relationships. In the decoder decoding stage, image reconstruction is completed at the high resolution layer by progressive upsampling, and the feature map is restored to the target high resolution output size to obtain the target facial profile correction prediction map.

[0050] In the above scheme, the local enhancement network adopts a pre-defined second end-to-end convolutional architecture. First, it extracts local feature maps from the input image through progressive downsampling. Then, it fuses these local feature maps with the initial facial profile correction prediction map output by the global generator network to obtain an initial fused feature map. After convolutional transformation to obtain the target fused feature map, it progressively upsamples at high-resolution levels to obtain the target facial profile correction prediction map. Thus, by extracting local texture and edge detail features from the image through downsampling and fusing them with global structural features, a coordinated expression of global layout and local details can be achieved. The high-resolution convolutional transformation and upsampling further refine high-frequency detail information, compensating for the insufficient detail expression of a single generator network at high resolution. Ultimately, it can generate a high-resolution correction prediction image that combines overall structural consistency and local texture realism, helping to improve the accuracy of postoperative facial profile correction prediction results.

[0051] In a preferred embodiment, the discriminator construction unit 402 includes: a sub-discriminator network for acquiring the resolution of an input image; if the resolution is less than or equal to a preset first resolution threshold, then the overall structural information of the input image is discriminated to generate a first discrimination result; if the resolution is greater than the preset first resolution threshold and less than a preset second resolution threshold, then the mesoscale structural features of the input image are discriminated to generate a second discrimination result; if the resolution is greater than or equal to the preset second resolution threshold, then the local texture and edge contours of the input image are discriminated to generate a third discrimination result.

[0052] For example, see Figure 5 , Figure 5 This is a schematic diagram of the discriminator structure of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown. Figure 5 As shown, in this embodiment, three sub-discriminator networks with the same network structure but independent parameters are used as examples for explanation. These three sub-discriminator networks are denoted as D1, D2 and D3. First, the input real image and the target facial profile correction prediction map output by the generator are subjected to multi-scale downsampling processing to construct input data at different resolution levels, and then input into the corresponding sub-discriminator networks respectively. Before proceeding, a low-resolution threshold is set as the first resolution threshold, and a high-resolution threshold is set as the second resolution threshold. The low-resolution and high-resolution thresholds can be set according to the actual application. In this example, sub-discriminator network D1 receives the low-resolution image with the largest downsampling ratio. The low-resolution image is the target facial profile correction prediction image with a resolution less than or equal to the preset first resolution threshold, and the input real image. The large effective receptive field of sub-discriminator network D1 can discriminate the overall structural information of the image, mainly used to constrain the rationality of the generated image in terms of global semantic layout, overall contour, and spatial structure, obtaining the first discrimination result. Sub-discriminator network D2 receives a medium-resolution image, used for... The structural features of the mesoscale resolution image are supervised. The mesoscale resolution image is the target facial profile correction prediction map with a resolution greater than a preset first resolution threshold and less than a preset second resolution threshold, and the input real image. This ensures the naturalness of the facial structure transition and the harmony of the overall proportion, resulting in the second discrimination result. The sub-discriminator network D3 receives a high-resolution image close to the target resolution. The high-resolution image is the target facial profile correction prediction map with a resolution greater than or equal to the preset second resolution threshold, and the input real image. The smaller receptive field of the sub-discriminator network D3 can more sensitively capture high-frequency detail information in the image, which is used to judge the authenticity of local texture, edge contours and fine structures, resulting in the third discrimination result.

[0053] In the above scheme, multiple sets of parameter-independent sub-discriminator networks are constructed. Images of different resolutions are input into the corresponding sub-discriminator networks through multi-scale downsampling, and the corresponding first, second, or third discrimination results are output. Thus, by utilizing the different receptive field ranges of the sub-discriminator networks at different resolutions, the authenticity of the generated images is verified from multiple levels, including global structure, mesoscale contours, and local details. This ensures that the images generated by the subsequent model conform to the features of real images in both overall structure and local details, helping to improve the accuracy of postoperative facial profile correction prediction results.

[0054] In a preferred embodiment, the model training module 104 includes a forward propagation unit 501, an optimization target construction unit 502, and a model parameter update unit 503; wherein: the forward propagation unit 501 is used to obtain a training dataset from the standard paired training dataset, input the training dataset into the adversarial network model to be trained, and obtain the first discrimination result, the second discrimination result, or the third discrimination result; the optimization target construction unit 502 is used to calculate and generate adversarial loss and feature matching loss based on the target loss function, according to the first discrimination result, the second discrimination result, or the third discrimination result, and perform a weighted combination of the adversarial loss and the feature matching loss to obtain the optimization target; the model parameter update unit 503 is used to alternately update the generator and discriminator in the adversarial network model to be trained according to the optimization target until the model parameters of the adversarial network model to be trained meet the preset training requirements, thereby obtaining the target adversarial network model.

[0055] For example, the standard paired training dataset is first randomly divided into training, validation, and test datasets in a 7:2:1 ratio. The batch size is set to 8 based on GPU memory limitations; the Adam optimizer is used, and the initial learning rate is set to 0.0002. An improved linear decay strategy is employed for learning rate scheduling: the initial learning rate is kept constant for the first 100 epochs to ensure sufficient model convergence and structural stability; in the subsequent 200 epochs, the learning rate is linearly decreased to 0. The entire training process lasts for 300 epochs. Model checkpoints are periodically saved during training, and the weight parameters that perform best on the validation set are finally selected for subsequent model training.

[0056] See Figure 6 , Figure 6 This is a schematic diagram of the prediction simulation process of a facial profile correction result prediction system according to a certain embodiment of the present invention; as shown below. Figure 6 As shown, paired image data from the training dataset is input into the adversarial network model to be trained. The input image is first processed by the generator to generate the corresponding predicted profile image. Then, the generated predicted image and the real target image are fed into the multi-scale discriminator. The discriminator first performs downsampling processing on the input image at different ratios to obtain multiple sets of image data at different resolution levels, and inputs them into sub-discriminator networks with independent parameters. After multi-layer convolutional feature extraction and normalization processing, each sub-discriminator network outputs the discrimination result of the corresponding level. In this embodiment, the low-resolution sub-discriminator network outputs the first discrimination result corresponding to the overall structure, the medium-resolution sub-discriminator network outputs the second discrimination result corresponding to the medium-scale structure, and the high-resolution sub-discriminator network outputs the third discrimination result corresponding to the local details. At the same time, each sub-discriminator network synchronously outputs multi-layer intermediate feature mapping data. Then, based on the target loss function framework, the generative adversarial loss and feature matching loss are calculated separately. The generative adversarial loss is calculated based on the true / false discrimination results output by each level of the sub-discriminator network, and is expressed as: ; The feature matching loss is based on the intermediate layer feature maps output by each sub-discriminator network, calculating the feature differences between the real target image and the generated predicted image in the multi-layer feature space. The feature matching loss is expressed as: ; In the formula, For input data, For label data, The total number of layers in the feature space. Data generated by the generator Let i be the total number of feature differences in the i-th layer. The features extracted from the input data at the i-th layer of the discriminator. To generate the features extracted from the i-th layer of the discriminator, The L1 distance symbol is used to measure the difference between two data points. The average value is calculated by summing up the total differences in characteristics; Then, corresponding weight coefficients are assigned to the two types of losses and weighted combinations are performed to form a unified overall optimization objective, expressed as: ; In the formula, These are the weighting coefficients; An adversarial training strategy with alternating updates is adopted to update the parameters of the generator network module and the discriminator network module respectively, so that the two gradually reach a dynamic equilibrium during the adversarial game. Training is terminated when the model loss tends to stabilize and the verification effect meets the preset training requirements. In this embodiment, the preset training requirements can be to meet the preset number of iterations. Finally, the network weight parameters that perform best on the verification subset are selected and solidified to obtain a target adversarial network model that can be directly used for actual prediction tasks.

[0057] In the above scheme, training data is first input into the model to be trained to obtain the discrimination results at each level. Then, adversarial loss and feature matching loss are calculated based on the target loss function. After weighted combination, the overall optimization target is obtained. Finally, the parameters of the generator and discriminator are alternately updated according to the optimization target until the model parameters meet the preset training requirements, thus obtaining the target adversarial network model. Therefore, multi-scale discrimination information is obtained through forward propagation, providing multi-dimensional basis for loss calculation. The composite loss function can simultaneously constrain the visual realism and feature-level consistency of the generated image. The alternating update training method enables adversarial game and collaborative optimization between the generator and discriminator, ultimately enabling the model to converge stably and learn accurate facial profile correction mapping relationships. This constructs a reliable target adversarial network model, thereby improving the accuracy of postoperative facial profile correction prediction results.

[0058] In a preferred embodiment, the result prediction module 105 includes a to-be-predicted image acquisition unit 601 and a prediction result output unit 602; wherein: the to-be-predicted image acquisition unit 601 is used to acquire a to-be-predicted facial profile image; the prediction result output unit 602 is used to input the to-be-predicted facial profile image into the target adversarial network model, perform feature extraction and multi-scale feature fusion on the to-be-predicted facial profile image through the generator in the target adversarial network model, generate a corresponding facial profile correction prediction image, and output the facial profile correction prediction image as the facial profile correction prediction result.

[0059] For example, Figure 7 This is a schematic diagram illustrating the prediction result output of a facial profile correction prediction system according to a certain embodiment of the present invention. Figure 7 As shown, real-time acquired preoperative facial profile images are used as the input for prediction. Following preprocessing rules identical to those used in the model training phase, the images undergo size calibration, resolution standardization, and pixel value normalization. The processed images are then transmitted to the target adversarial network (TAN) model. The TAN model extracts global structural information and local feature information from the images through its internal hierarchical network structure. Global structural information includes macroscopic features such as the overall facial contour proportions and maxillofacial skeletal layout, while local feature information covers microscopic details such as the mandibular border line, chin shape, and soft tissue texture transitions. The generator of the TAN model then automatically generates a high-resolution facial profile correction prediction image based on the feature mapping association rules accumulated during training, outputting the facial profile correction prediction result.

[0060] In the above scheme, a facial profile image to be predicted is first acquired. Then, the generator in the target adversarial network model extracts features and fuses multi-scale features from the facial profile image. The trained generator automatically generates a facial profile correction prediction image, which is then output as the correction prediction result. Thus, by automatically extracting multi-dimensional features from the image to be predicted using the target model, and automatically generating the corresponding correction result based on the mapping relationship between preoperative facial structure features and postoperative appearance changes learned during network training, pixel-level editing is eliminated, ensuring that the prediction result follows the objective correction rules learned by the model. This helps improve the accuracy of postoperative facial profile correction prediction results.

[0061] This invention also provides a facial profile correction result prediction device, the device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor applies the facial profile correction result prediction system when executing the computer program.

[0062] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A facial profile correction result prediction system, characterized in that, It includes a dataset construction module, a dataset preprocessing module, a model construction module, a model training module, and a result prediction module; among which: The dataset construction module is used to perform image correction on the pre-collected initial facial profile image using a preset correction algorithm to obtain the target facial profile image, and to construct a paired dataset based on the mapping relationship between the initial facial profile image and the target facial profile image. The dataset preprocessing module is used to obtain a standard pairing training dataset by randomly cropping each pairing data in the pairing dataset. The model building module is used to cascade a global generator network and a local enhancement network to build a generator, build a discriminator based on several sets of parameter-independent sub-discriminators, and build an adversarial network model to be trained based on the generator and the discriminator. The model training module is used to input the standard paired training dataset into the adversarial network model to be trained, and iteratively update the model parameters of the adversarial network model to be trained through a preset training strategy until the model parameters meet the preset training requirements, thereby obtaining the target adversarial network model. The result prediction module is used to input the facial profile image to be predicted into the target adversarial network model, and output the facial profile correction prediction result through the target adversarial network model.

2. The facial profile correction result prediction system as described in claim 1, characterized in that, The dataset construction module includes a validity screening unit, an image correction unit, and a data pairing unit; wherein: The validity screening unit is used to collect initial facial profile images and remove initial facial profile images that are occluded, blurred, or distorted based on preset validity screening indicators to obtain valid initial facial profile images. The image correction unit is used to perform contour correction and texture restoration on the deformed areas in the effective initial facial profile image based on a preset correction algorithm, so as to obtain a target facial profile image that is mapped one-to-one with the effective initial facial profile image. The data pairing unit is used to pair and bind each pair of corresponding valid initial facial profile images and target facial profile images to form a paired dataset.

3. The facial profile correction result prediction system as described in claim 2, characterized in that, The dataset preprocessing module includes an image cropping unit, an image preprocessing unit, and a data integration unit; wherein: The image cropping unit is used to randomly crop the effective initial facial profile image and the target facial profile image in each set of initial paired data based on a preset ratio and in a preset cropping area to obtain several local images. The image preprocessing unit is used to restore the size of several local images to be consistent with the size of the target facial profile image by using a preset magnification algorithm, and to perform resolution standardization and pixel value normalization on the several local images after size restoration to obtain several sets of target pairing data. The data integration unit is used to integrate each set of initial pairing data and several sets of target pairing data into a standard pairing training dataset.

4. The facial profile correction result prediction system as described in claim 1, characterized in that, The model building module includes a generator building unit, a discriminator building unit, and a trainable adversarial network model building unit; The generator building unit is used to build a global generation network based on a preset first end-to-end convolutional architecture and a preset first resolution level, and to build a local enhancement network based on a preset second end-to-end convolutional architecture and a preset second resolution level. The global generation network and the local enhancement network are cascaded to obtain a generator. The resolution of the preset first resolution level is smaller than the resolution of the preset second resolution level. The discriminator construction unit is used to set independent and different network configuration parameters for each of the initial sub-discriminators based on several sets of initial sub-discriminators with the same network structure, to obtain several sets of sub-discriminators, and to combine the several sets of sub-discriminators into a discriminator; The adversarial network model building unit is used to construct a target loss function based on the generator and the discriminator, and to construct an adversarial network model to be trained based on the target loss function.

5. The facial profile correction result prediction system as described in claim 4, characterized in that, The generator building unit includes: The global generation network is used to downsample the input image step by step according to a preset first end-to-end convolutional architecture to obtain an initial global semantic feature map, perform residual propagation on the initial global semantic feature map based on a preset nonlinear activation function to obtain a target global semantic feature map, and perform upsampling on the target global semantic feature map step by step based on a preset first resolution level to obtain an initial facial profile correction prediction map.

6. The facial profile correction result prediction system as described in claim 5, characterized in that, The generator building unit includes: The local enhancement network is used to downsample the input image step by step according to a preset second end-to-end convolutional architecture, extract local feature maps, fuse the local feature maps and the initial facial profile correction prediction map to obtain an initial fused feature map, perform convolutional transformation on the initial fused feature map to obtain a target fused feature map, and perform step-by-step upsampling on the target fused feature map based on a preset second resolution level to obtain a target facial profile correction prediction map.

7. The facial profile correction result prediction system as described in claim 6, characterized in that, The discriminator construction unit includes: The sub-discriminator network is used to obtain the resolution of the input image. If the resolution is less than or equal to a preset first resolution threshold, the overall structural information of the input image is discriminated to generate a first discrimination result. If the resolution is greater than the preset first resolution threshold and less than a preset second resolution threshold, the mesoscale structural features of the input image are discriminated to generate a second discrimination result. If the resolution is greater than or equal to the preset second resolution threshold, the local texture and edge contours of the input image are discriminated to generate a third discrimination result.

8. The facial profile correction result prediction system as described in claim 7, characterized in that, The model training module includes a forward propagation unit, an optimization target construction unit, and a model parameter update unit; wherein: The forward propagation unit is used to obtain a training dataset from the standard paired training dataset, input the training dataset into the adversarial network model to be trained, and obtain the first discrimination result, the second discrimination result, or the third discrimination result; The optimization target construction unit is used to calculate and generate adversarial loss and feature matching loss based on the target loss function, according to the first discrimination result, the second discrimination result or the third discrimination result, and to obtain the optimization target by weighting and combining the adversarial loss and the feature matching loss. The model parameter update unit is used to alternately update the generator and discriminator in the adversarial network model to be trained according to the optimization objective, until the model parameters of the adversarial network model to be trained meet the preset training requirements, thereby obtaining the target adversarial network model.

9. The facial profile correction result prediction system as described in claim 3, characterized in that, The result prediction module includes a to-be-predicted image acquisition unit and a prediction result output unit; wherein: The image acquisition unit is used to acquire the facial profile image to be predicted. The prediction result output unit is used to input the facial profile image to be predicted into the target adversarial network model, and to perform feature extraction and multi-scale feature fusion on the facial profile image to be predicted through the generator in the target adversarial network model to generate a corresponding facial profile correction prediction image, and output the facial profile correction prediction image as the facial profile correction prediction result.

10. A facial profile correction result prediction device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, applies a facial profile correction result prediction system as described in any one of claims 1 to 9.

Images