Large-angle facial expression recognition method and system based on multi-view evidence fusion

Abstract

The invention discloses a multi-view evidence fusion large-angle facial expression recognition method and system, and the method mainly comprises the steps: directly transmitting a read image into a face detection model based on a YOLOv8 architecture, returning all detected face targets in the image, extracting key information, and calculating the in-plane rotation angle of a face; when the absolute value of the rotation angle exceeds a threshold value, activating the GAN model to generate a view 2, and cutting a face region from the original image by using the returned face target all the time regardless of the rotation angle to serve as a view 1; extracting the view 1 and the view 2 in parallel to obtain two groups of independent evidence vectors, and calculating subjective opinions corresponding to the evidence vectors; if the two views are processed, evidence fusion is started, otherwise fusion is not needed, an index of the maximum value is searched in a final belief vector to determine the most credible expression category, and final uncertain expressions are output together. Compared with any single view or simple fusion method, the method provided by the invention can obtain more accurate and more stable recognition performance under various attitude angles.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and computer vision technology, specifically relating to facial expression recognition methods and systems. Background Technology

[0002] Facial expressions are key nonverbal cues for humans to communicate emotional states. Automated facial expression recognition (FER) technology has broad application potential in fields such as human-computer interaction, mental health assessment, autonomous driving, and market analysis. With the development of deep learning, especially convolutional neural networks (CNNs), FER technology has made significant progress under controlled conditions (mainly frontal or near-frontal facial images), and its recognition accuracy continues to improve.

[0003] However, real-world applications are far more complex than laboratory environments. Variations in head pose are a major challenge. When a face is significantly angulated relative to the camera, the facial geometry undergoes significant non-linear changes, and partial facial occlusion may occur. These changes lead to a large discrepancy between the visual features extracted from the image for expression classification and the frontal features learned by the model during training. Therefore, existing FER models, even high-performing deep models, typically experience a sharp decline in generalization ability when dealing with faces at large angles.

[0004] While Generative Adversarial Networks (GANs) offer an effective approach to this problem by "frontalizing" large-angle facial images for pose correction, this approach also has limitations. GAN-generated images may introduce imperceptible artifacts or lose subtle facial muscle changes crucial for expression classification. A simple binary choice between the "original image" and the "generated image" may discard useful information or introduce unwanted noise. A deeper challenge lies in the uncertainty inherent in any single source of information. Traditional deep learning models typically use the Softmax function to output classification confidence, but this confidence is often "overconfident," potentially outputting extremely high scores even when predictions are incorrect, failing to accurately reflect the model's predictive reliability. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to overcome the limitations of existing single-view models and simple fusion strategies in handling large-angle facial expression recognition.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a large-angle facial expression recognition method based on multi-view evidence fusion, which includes the following steps after system environment configuration and initialization: The read image is directly fed into the initialized face detection model based on the YOLOv8 architecture for inference, and all detected face targets in the image are returned. For each face target, key information is extracted, and the in-plane rotation angle of the face is calculated using the extracted key information. The system compares the absolute value of the rotation angle with a preset threshold. When the absolute value of the rotation angle exceeds the threshold, the system determines that the current face is in a large-angle pose and activates the GAN model to generate a frontalized face image that retains identity information and core facial features, which serves as the visual representation. Figure 2 If the rotation angle does not exceed the threshold, no view will be generated. Figure 2 Regardless of the rotation angle, the returned face target is always used to crop the face region from the original image as the viewfinder. Figure 1 ; View Figure 1 Heshi Figure 2 Two independent evidence vectors are extracted in parallel. For each evidence vector, the corresponding subjective opinion is calculated. If two views are processed, evidence fusion is initiated, taking the two sets of subjective opinions as input, and calculating the final fused belief vector and fused uncertainty scalar. If only one view is processed, fusion is not required, and its own belief vector and uncertainty scalar are the final results. The index of the maximum value in the final belief vector is found to determine the most credible expression category, and the final uncertainty is output together.

[0007] As a further optimized technical solution, for each detected face target, two key pieces of information are extracted: bounding box coordinates bbox (x1, y1, x2, y2): which defines the precise position of the face in the image; and five facial key points: including the (x, y) coordinates of the left eye, right eye, tip of the nose, left corner of the mouth, and right corner of the mouth. Pose quantization: Using the extracted coordinates of the left and right eye key points, the in-plane rotation angle of the face is calculated by calculating the angle between the line connecting the two points and the horizontal axis.

[0008] As a further optimized technical solution, the in-plane rotation angle of the face is calculated by calculating the angle between the line connecting two points and the horizontal axis, specifically including: Given left eye coordinates ( , ) and right eye coordinates ( , ) Vertical difference: = Horizontal difference: = Rotation angle (radians):

[0009] Rotation angle (degrees): = (180 / ).

[0010] As a further optimized technical solution, [the following is a description of the solution]. Figure 2 The generation process is as follows: First, the view... Figure 1 The image is preprocessed to adapt to the GAN model: it is converted from OpenCV's BGR format to Pillow's RGB format, the image size is uniformly adjusted to 128x128 pixels, then converted to a PyTorch tensor, and the pixel values ​​are normalized to the [-1,1] interval. The processed tensor is fed into the GAN model for inference. The output tensor of the GAN model is then reverse-processed, i.e., de-normalized to the [0,255] interval and converted back to OpenCV image format to obtain a frontalized face image, which is used as the viewfinder. Figure 2 .

[0011] As a further optimized technical solution, it will be considered Figure 1 Heshi Figure 2 Parallel extraction yields two independent sets of evidence vectors, including: view Figure 1 Heshi Figure 2 The images were converted to grayscale, and their dimensions were uniformly adjusted to 64x64 pixels. The pixel values ​​were converted to floating-point numbers and normalized to the [0,1] interval by dividing by 255.0. Then, they were fed in parallel into the modified evidence extraction network to obtain two independent 7-dimensional evidence vectors. and .

[0012] As a further optimized technical solution, the evidence extraction network adopts the Mini-XCEPTION network architecture as the basic model. After loading a pre-trained expression classification model based on the Mini-XCEPTION architecture, its final Softmax activation layer and classification layer are removed, and a new fully connected layer is connected on top of it. The number of output units of this layer is the predefined number of expression categories, and Softplus is used as its activation function.

[0013] As a further optimized technical solution, for each obtained evidence vector, the corresponding subjective opinion is calculated, including: for each obtained evidence vector, applying the subjective logic transformation formula to calculate its corresponding belief vector and uncertainty scalar; for any view m (m=1,2), its subjective opinion... Quality of Belief and uncertainty quality The composition and subjective logical transformation formula are as follows: (1) Calculate the Dirichlet distribution parameters: +1, where It is an evidence vector The value of the k-th element; (2) Calculate the Dirichlet intensity: ; (3) Calculate the quality of beliefs: = = ; (4) Calculate the uncertainty mass: = , where K is the total number of categories.

[0014] As a further optimized technical solution, the two sets of subjective opinions are used as input to calculate the final fused belief vector and fused uncertainty scalar, including: the two sets of subjective opinions { , }and{ , As input, the Dempster-Shafer combination rule is invoked, and the final fusion belief vector is obtained by calculating the conflict degree, fusion belief, and fusion uncertainty. and fusion uncertainty scalar .

[0015] As a further optimized technical solution, the process of invoking the Dempster-Shafer combination rule and calculating the conflict degree, fusion belief, and fusion uncertainty is as follows: Calculate the degree of conflict: ,in Indicates vision Figure 1 Beliefs in category i Indicates vision Figure 2 Beliefs about category j; Computational fusion of beliefs: (1 / (1 C))*( + ,in For the beliefs about category k after fusion, and respectively view Figure 1 Heshi Figure 2 Beliefs about category k and Due to the uncertainty of each view; Uncertainties after computational fusion: .

[0016] This invention also provides a large-angle facial expression recognition system based on multi-view evidence fusion, comprising the following modules: System environment configuration and initialization module; The face detection and pose quantization module is used to directly feed the read image into the initialized face detection model based on the YOLOv8 architecture for inference, and return all detected face targets in the image. For each face target, key information is extracted, and the in-plane rotation angle of the face is calculated using the extracted key information. The conditional multi-view generation module compares the absolute value of the rotation angle with a preset threshold. When the absolute value of the rotation angle exceeds the threshold, the system determines that the current face is in a large-angle pose and activates the GAN model to generate a frontalized face image that retains identity information and core facial features, which serves as the view. Figure 2 If the rotation angle does not exceed the threshold, no view will be generated. Figure 2 Regardless of the rotation angle, the returned face target is always used to crop the face region from the original image as the viewfinder. Figure 1 ; The parallel evidence extraction and opinion transformation module is used to transform visual evidence into opinions. Figure 1 Heshi Figure 2 Two independent evidence vectors are extracted in parallel. For each evidence vector, the corresponding subjective opinion is calculated. The evidence fusion and credible decision output module is used to determine whether evidence fusion is initiated if two views have been processed. The two sets of subjective opinions are taken as input, and the final fused belief vector and fused uncertainty scalar are calculated. If only one view has been processed, no fusion is required, and its own belief vector and uncertainty scalar are the final results. The index of the maximum value in the final belief vector is found to determine the most credible expression category, and the final uncertainty is output together.

[0017] The advantages of this invention are: 1. Significantly improves accuracy and robustness: By intelligently fusing the real information of the original view and the pose correction information of the GAN view, this invention can complement the advantages of both, and achieve more accurate and stable recognition performance than any single view or simple fusion method under various pose angles, especially large angles.

[0018] 2. Ensuring Credibility and Reliability in Decision Making: The greatest innovation of this invention lies in the introduction of an uncertainty quantification mechanism. The system can clearly identify difficult samples caused by insufficient evidence (such as blurry images or extreme lighting) or conflicting evidence (such as two views giving drastically different judgments), and warn against them with a high uncertainty value, fundamentally avoiding the "overconfidence" defect of traditional classifiers.

[0019] 3. Theoretical Completeness and Adaptability of the Fusion Strategy: Unlike heuristic fusion methods (such as weighted averaging), the Dempster-Shafer evidence theory used in this invention is a framework with a solid mathematical foundation. It can dynamically and adaptively aggregate evidence based on the certainty of each view's judgment on a specific sample, without the need for manual weight setting. Attached Figure Description

[0020] Figure 1 This is a flowchart of a large-angle facial expression recognition method based on multi-view evidence fusion according to an embodiment of the present invention; Figure 2 This is a diagram of the generated adversarial network structure in the embodiment; Figure 3 This is a schematic diagram illustrating the effect of generating adversarial networks in the embodiment; Figure 4 This is a diagram of the emotion recognition network structure in the embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The core of the technical solution proposed in this invention is to construct an end-to-end facial expression recognition pipeline based on multi-view evidence theory. Its overall architecture is designed as follows: the input image first passes through a preprocessing module to generate multiple views, then each view enters the evidence extraction network in parallel. The output evidence is converted into subjective opinions containing beliefs and uncertainties, and finally, a credible decision is obtained through the DS evidence fusion module.

[0023] See Figure 1 As shown, the large-angle credible facial expression recognition method based on multi-view evidence fusion of the present invention includes the following steps: S1. System environment configuration and initialization, including: Hardware environment: This solution is deployed on a standard high-performance computer equipped with a graphics processing unit (GPU) that supports the CUDA architecture to meet the computing power requirements of deep learning model inference.

[0024] Software environment: Operating system is Ubuntu 18.04 or a compatible version. Programming language is Python 3.x. The core functionality of the system depends on the following third-party libraries: Ultralytics: Used for high-performance, high-precision face detection.

[0025] PyTorch: Used for loading and executing GAN-based face frontalization models.

[0026] TensorFlow (and Keras): Used to load and execute basic facial expression classification models and modify their network structure.

[0027] OpenCV-Python: Used for core image processing tasks such as image reading, color space conversion, resizing, and result visualization.

[0028] NumPy: Used for efficient multidimensional array operations, especially in attitude angle calculations and the implementation of evidence theory formulas.

[0029] Pillow (PIL): Used for image format conversion between different image processing libraries (such as OpenCV and PyTorch).

[0030] Model Loading and Preparation: Upon system startup or first run, the following initialization operations are performed: three core pre-trained models are loaded into memory: a face detection model based on the YOLOv8 architecture, a pre-trained GAN (Generative Adversarial Network) model specifically for face frontalization, and a pre-trained expression classification model based on the Mini-XCEPTION architecture. Key online modifications are then made to the loaded basic expression classification model (based on the mature Mini-XCEPTION architecture). Specifically, the Softmax activation layer and the final classification layer at the end of the original network structure are removed, and a new fully connected (Dense) layer is added to the top of the network. The number of output neurons in this fully connected layer is set to the total number of expression categories (7 in this scheme), and Softplus is used as its activation function. This modified network no longer outputs probabilities but instead outputs non-negative "evidence" vectors required for subsequent evidence theory calculations, which are used as a unified "evidence extractor" in subsequent processes.

[0031] S2, Face Detection and Pose Quantization Read the input image; Face localization: The read BGR image is directly fed into the initialized YOLOv8-based face detection model for inference. The YOLOv8-based face detection model will return all detected face targets in the image.

[0032] Information extraction: For each detected face target, two key pieces of information are extracted: bounding box coordinates bbox(x1,y1,x2,y2): which defines the precise location of the face in the image; and 5 facial key points: including the (x,y) coordinates of the left eye, right eye, tip of the nose, left corner of the mouth, and right corner of the mouth.

[0033] Pose quantization: Using the extracted coordinates of the left and right eye key points, the in-plane rotation angle of the face is calculated by calculating the angle between the line connecting the two points and the horizontal axis. The specific calculation formula is as follows: Given left eye coordinates ( , ) and right eye coordinates ( , ) Vertical difference: = Horizontal difference: = Rotation angle (radians):

[0034] Rotation angle (degrees): = (180 / ).

[0035] S3, Conditional Multi-View Generation This step prepares the input view for the subsequent evidence fusion module based on the results of S2 pose quantization. The rotation angle (in degrees) calculated by S2 is then used. The absolute value is compared with a preset threshold (e.g., 30 degrees).

[0036] For vision Figure 1 (Original Face View): Regardless of the rotation angle, the system always uses the face target returned by the face detection model based on the YOLOv8 architecture to crop out the face region (ROI) from the original image as the view. Figure 1 .

[0037] For vision Figure 2 (Front-facing face view): The system has a built-in angle threshold (e.g., 30 degrees). When the absolute value of the calculated rotation angle exceeds this threshold, the system determines that the current face is in a large-angle pose and activates a pre-trained GAN model specifically designed for front-facing face mapping (its network structure is as follows). Figure 2 As shown, the positive effect is compared to... Figure 3 As shown), a frontalized facial image is generated that retains identity information and core facial features, serving as the visual representation. Figure 2 If the rotation angle does not exceed the threshold, no view will be generated. Figure 2Subsequent processes will be based solely on video. Figure 1 Proceed. (View) Figure 2 The generation process is as follows: First, the view... Figure 1 The image is preprocessed to fit the GAN model: it is converted from OpenCV's BGR format to Pillow's RGB format, the image size is uniformly adjusted to 128x128 pixels, then converted to a PyTorch tensor, and the pixel values ​​are normalized to the [-1,1] interval. The processed tensor is fed into the GAN model for inference. The output tensor of the GAN model is then reverse-processed (inversely normalized to the [0,255] interval and converted back to OpenCV image format) to obtain a frontalized face image, which is used as the view frustum. Figure 2 .

[0038] S4. Parallel Evidence Extraction and Opinion Transformation 1) Parallel Evidence Extraction: For each view, perform the following operations: Preprocessing: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Figure 1 Heshi Figure 2 The image (if present) is converted to grayscale, its size is uniformly adjusted to 64x64 pixels, and the pixel values ​​are converted to floating-point numbers and normalized to the [0,1] interval by dividing by 255.0. These values ​​are then fed in parallel into the modified evidence extraction network to obtain two independent 7-dimensional evidence vectors. and .

[0039] This invention uses a specially modified deep convolutional neural network as a unified evidence extraction network. The scheme employs the Mini-XCEPTION network architecture as the base model (Arriaga, O., Valdenegro-Toro, M., & Plöger, P., 2017), its structure as follows: Figure 4As shown, the network mainly consists of an input layer, regular convolutional layers, and residual blocks. The specific process is as follows: the input image first passes through two convolutional layers (Conv2D) and batch normalization (BatchNorm); then it enters four consecutive depthwise separable convolutional residual blocks, each containing a 1x1 convolution, a depthwise separable convolution (Sep-Conv2D), max pooling, and a residual connection (Add); finally, it passes through convolutional layers and a global average pooling layer. This network model performs excellently in facial expression recognition tasks. After loading a pre-trained facial expression classification model based on the Mini-XCEPTION architecture, its final Softmax activation layer and classification layer are removed. A new fully connected (Dense) layer is connected on top, with the number of output units equal to the predefined number of facial expression categories (7 in this scheme), and Softplus is used as its activation function. The Softplus function guarantees that all output values ​​are non-negative, making them suitable as evidence for subsequent evidence theory calculations, rather than probabilities in the traditional sense.

[0040] 2) Subjective Opinion Transformation: For each obtained evidence vector E, apply the subjective logic transformation formula to accurately calculate its corresponding belief vector b and uncertainty scalar u. For any view m (m=1,2), its subjective opinion... Quality of Belief and uncertainty quality The composition and subjective logical transformation formula are as follows: (1) Calculate the Dirichlet distribution parameters: +1, where It is an evidence vector The value of the k-th element; (2) Calculate the Dirichlet intensity: ; (3) Calculate the quality of beliefs: = = ; (4) Calculate the uncertainty mass: = , where K is the total number of categories (7 in this case).

[0041] S5, Evidence Fusion and Credible Decision Output 1) Fusion Decision: The system checks how many views were processed in step S4. If two views (views) were processed... Figure 1 Heshi Figure 2 If ), then evidence fusion is initiated, combining the two sets of subjective opinions { , }and{ , As input, the Dempster-Shafer combination rule is invoked, and the final fusion belief vector is obtained by calculating the conflict degree, fusion belief, and fusion uncertainty. and fusion uncertainty scalar If only one view is processed (view ... Figure 1 If the belief vector and uncertainty scalar are independent, then fusion is unnecessary. The final result is its own belief vector and uncertainty scalar. = and = .

[0042] The process of invoking the Dempster-Shafer combination rule and calculating the degree of conflict, fusion belief, and fusion uncertainty is as follows: Calculate the degree of conflict: ;(in Indicates vision Figure 1 Beliefs in category i Indicates vision Figure 2 (Beliefs about category j) Computational fusion of beliefs: (1 / (1 C))*( + ;(in For the beliefs about category k after fusion, and respectively view Figure 1 Heshi Figure 2 Beliefs about category k and (Due to the uncertainty of each view) Uncertainties after computational fusion: .

[0043] 2) Decision output: Facial expression classification: through the final belief vector The index of the maximum value is used to determine the most believable expression category (e.g., "happy", "sad", etc.).

[0044] Credibility assessment: addressing the ultimate uncertainty This value is output along with the classification results as a clear, quantifiable credibility indicator. The closer this value is to 0, the more credible the decision; the higher the value, the less credible the decision, indicating insufficient or conflicting evidence. The system can set an uncertainty threshold (e.g., 0.5). When this value is exceeded, the final output can be marked as "uncertain" to avoid outputting potentially incorrect conclusions under low-certainty conditions.

[0045] The present invention also provides a system corresponding to the above-mentioned large-angle facial expression recognition method based on multi-view evidence fusion, comprising the following modules: System environment configuration and initialization module; The face detection and pose quantization module is used to directly feed the read image into the initialized face detection model based on the YOLOv8 architecture for inference, and return all detected face targets in the image. For each face target, key information is extracted, and the in-plane rotation angle of the face is calculated using the extracted key information. The conditional multi-view generation module compares the absolute value of the rotation angle with a preset threshold. When the absolute value of the rotation angle exceeds the threshold, the system determines that the current face is in a large-angle pose and activates the GAN model to generate a frontalized face image that retains identity information and core facial features, which serves as the view. Figure 2 If the rotation angle does not exceed the threshold, no view will be generated. Figure 2 Regardless of the rotation angle, the returned face target is always used to crop the face region from the original image as the viewfinder. Figure 1 ; The parallel evidence extraction and opinion transformation module is used to transform visual evidence into opinions. Figure 1 Heshi Figure 2 Two independent evidence vectors are extracted in parallel. For each evidence vector, the corresponding subjective opinion is calculated. The evidence fusion and credible decision output module is used to determine whether evidence fusion is initiated if two views have been processed. The two sets of subjective opinions are taken as input, and the final fused belief vector and fused uncertainty scalar are calculated. If only one view has been processed, no fusion is required, and its own belief vector and uncertainty scalar are the final results. The index of the maximum value in the final belief vector is found to determine the most credible expression category, and the final uncertainty is output together.

[0046] The execution programs of each module in this system are the same as the large-angle facial expression recognition method described above, which integrates multi-view evidence.

[0047] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for large-angle facial expression recognition using multi-view evidence fusion, characterized in that: After system environment configuration and initialization, the following steps are included: The read image is directly fed into the initialized face detection model based on the YOLOv8 architecture for inference, and all detected face targets in the image are returned. For each face target, key information is extracted, and the in-plane rotation angle of the face is calculated using the extracted key information. The absolute value of the rotation angle is compared with a preset threshold. When the absolute value of the rotation angle exceeds the threshold, the system determines that the current face is in a large-angle pose and activates the GAN model to generate a frontal face image that retains identity information and core expression features as View 2. If the rotation angle does not exceed the threshold, View 2 is not generated. Regardless of the rotation angle, the face region is always cropped from the original image using the returned face target as View 1. Two independent evidence vectors are obtained by extracting View 1 and View 2 in parallel. For each obtained evidence vector, the corresponding subjective opinion is calculated. If two views are processed, evidence fusion is initiated, taking the two sets of subjective opinions as input, and calculating the final fused belief vector and fused uncertainty scalar. If only one view is processed, fusion is not required, and its own belief vector and uncertainty scalar are the final results. The index of the maximum value in the final belief vector is found to determine the most credible expression category, and the final uncertainty is output together.

2. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 1, characterized in that: For each detected face target, two key pieces of information are extracted: bounding box coordinates (bbox (x1, y1, x2, y2): defining the precise location of the face in the image; and five facial key points: including the (x, y) coordinates of the left eye, right eye, tip of the nose, left corner of the mouth, and right corner of the mouth. Pose quantization: Using the extracted coordinates of the left and right eye key points, the in-plane rotation angle of the face is calculated by calculating the angle between the line connecting the two points and the horizontal axis.

3. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 2, characterized in that: The in-plane rotation angle of a human face is calculated by measuring the angle between the line connecting two points and the horizontal axis. Specifically, this includes: Given left eye coordinates ( , ) and right eye coordinates ( , ) Vertical difference: = Horizontal difference: = Rotation angle (radians): Rotation angle (degrees): = (180 / ).

4. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 1, characterized in that: The specific process for generating View 2 is as follows: First, the image of View 1 is preprocessed to adapt to the GAN model: it is converted from OpenCV's BGR format to Pillow's RGB format, the image size is uniformly adjusted to 128x128 pixels, then converted to a PyTorch tensor, and the pixel values ​​are normalized to the [-1,1] interval. The processed tensor is fed into the GAN model for inference. The output tensor of the GAN model is reverse-processed, that is, it is de-normalized to the [0,255] interval and converted back to the OpenCV image format to obtain a frontalized face image, which is used as View 2.

5. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 1, characterized in that: The parallel extraction of two independent evidence vectors from View 1 and View 2 involves: converting the images of View 1 and View 2 to grayscale, uniformly adjusting their size to 64x64 pixels, converting pixel values ​​to floating-point numbers and normalizing them to the [0,1] interval by dividing by 255.0, and then feeding them in parallel into the modified evidence extraction network to obtain two independent 7-dimensional evidence vectors. and .

6. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 5, characterized in that: The evidence extraction network uses the Mini-XCEPTION network architecture as the basic model. After loading a pre-trained facial expression classification model based on the Mini-XCEPTION architecture, its final Softmax activation layer and classification layer are removed, and a new fully connected layer is connected on top of it. The number of output units of this layer is the predefined number of facial expression categories, and Softplus is used as its activation function.

7. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 1, characterized in that: For each obtained evidence vector, the corresponding subjective opinion is calculated, including: for each obtained evidence vector, applying the subjective logic transformation formula to calculate its corresponding belief vector and uncertainty scalar; for any view m (m=1,2), its subjective opinion... Quality of Belief and uncertainty quality The composition and subjective logical transformation formula are as follows: (1) Calculate the Dirichlet distribution parameters: +1, where It is an evidence vector The value of the k-th element; (2) Calculate the Dirichlet intensity: ; (3) Calculate the quality of beliefs: = = ; (4) Calculate the uncertainty mass: = , where K is the total number of categories.

8. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 1, characterized in that: Using two sets of subjective opinions as input, the final fused belief vector and fused uncertainty scalar are calculated, including: the two sets of subjective opinions { , }and{ , As input, the Dempster-Shafer combination rule is invoked, and the final fusion belief vector is obtained by calculating the conflict degree, fusion belief, and fusion uncertainty. and fusion uncertainty scalar .

9. The large-angle facial expression recognition method based on multi-view evidence fusion as described in claim 8, characterized in that: in, The process of applying the Dempster-Shafer combination rule to calculate the degree of conflict, fusion belief, and fusion uncertainty is as follows: Calculate the degree of conflict: ,in This represents view 1's belief about category i. This represents view 2's belief about category j; Computational fusion of beliefs: (1 / (1 C))*( + ,in For the beliefs about category k after fusion, and The beliefs of View 1 and View 2 regarding category k are respectively. and Due to the uncertainty of each view; Uncertainties after computational fusion: .

10. A large-angle facial expression recognition system based on multi-view evidence fusion, characterized in that: Includes the following modules: System environment configuration and initialization module; The face detection and pose quantization module is used to directly feed the read image into the initialized face detection model based on the YOLOv8 architecture for inference, and return all detected face targets in the image. For each face target, key information is extracted, and the in-plane rotation angle of the face is calculated using the extracted key information. The conditional multi-view generation module compares the absolute value of the rotation angle with a preset threshold. When the absolute value of the rotation angle exceeds the threshold, the system determines that the current face is in a large-angle pose and activates the GAN model to generate a frontalized face image that retains identity information and core expression features as View 2. If the rotation angle does not exceed the threshold, View 2 is not generated. Regardless of the rotation angle, the returned face target is always used to crop the face region from the original image as View 1. The parallel evidence extraction and opinion transformation module is used to extract two independent sets of evidence vectors from View 1 and View 2 in parallel, and to calculate the corresponding subjective opinion for each obtained evidence vector. The evidence fusion and credible decision output module is used to determine whether evidence fusion is initiated if two views have been processed. The two sets of subjective opinions are taken as input, and the final fused belief vector and fused uncertainty scalar are calculated. If only one view has been processed, no fusion is required, and its own belief vector and uncertainty scalar are the final results. The index of the maximum value in the final belief vector is found to determine the most credible expression category, and the final uncertainty is output together.

Images