Efficient emotion recognition method and system

Abstract

The invention discloses a teenager emotion efficient identification method and system. The invention provides a solution based on an improved YOLO11 model (YOLO11-QX) aiming at lightweight and precision balance challenges of an existing emotion recognition model during complex facial expression capture of teenagers and resource-constrained equipment deployment. The main improvement comprises the following steps: introducing an SOCA module into a backbone network, and enhancing the fineness and robustness of facial feature extraction; and a DIOU loss function is adopted, so that more accurate bounding box regression optimization is realized, and the convergence efficiency of model training is improved. And through ONNX conversion and INT8 quantification processing, lightweight deployment is realized. Through the improvement, the YOLO11-QX model maintains the lightweight performance, and at the same time, the teenager emotion recognition accuracy is significantly improved. The system can be integrated in equipment such as monitoring and door lock cameras, a lightweight psychological emotion recognition network of scenes such as schools, communities and families can be independently deployed or constructed, and reliable data support is provided for early psychological intervention of teenagers.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence, computer vision and deep learning technology, and in particular relates to an emotion recognition method and system for adolescents. Background Technology

[0002] Emotion recognition technology, a hot research topic in the field of artificial intelligence, has shown broad application prospects in various scenarios such as human-computer interaction, mental health assessment, and intelligent monitoring. Especially against the backdrop of increasing attention to adolescent mental health, efficient and accurate methods for recognizing adolescent emotions are of great significance for timely detection of psychological problems and providing early intervention.

[0003] However, existing emotion recognition methods still face many challenges in practical applications. On the one hand, models designed for general adult emotion recognition often suffer from inaccurate feature capture and insufficient robustness when dealing with the complex facial expressions unique to adolescents. Emotional expressions during adolescence are subtle and changeable, and existing models may struggle to effectively distinguish the subtle differences between various emotions, leading to limited recognition accuracy. On the other hand, most emotion recognition models, in pursuit of high accuracy, typically employ deep learning network structures with large parameter counts and high computational complexity. This results in high computational resource consumption and slow inference speeds when deployed on edge devices or resource-constrained terminals (such as camera modules and smart lock devices), making it difficult to meet the practical requirements of real-time performance and lightweight design. This is particularly pronounced in scenarios requiring large-scale, distributed deployment, such as schools, communities, or homes, limiting their widespread application.

[0004] To address these challenges, current research focuses on improving the performance and deployment efficiency of emotion recognition models by optimizing model structure and introducing lightweight techniques. For example:

[0005] Early emotion recognition methods largely relied on hand-designed feature extractors, such as Local Binary Patterns (LBP) and Gabor features, combined with Support Vector Machines (SVM) or neural networks for classification. While these methods had relatively low computational cost, their feature extraction capabilities were limited, and they lacked robustness to complex backgrounds and facial expression changes.

[0006] In recent years, with the development of deep learning technology, convolutional neural networks (CNNs) have made significant progress in the field of facial expression recognition. For example, models based on classic backbone networks such as VGG, ResNet, and Inception can automatically learn high-level semantic features from images, thereby improving recognition accuracy. However, these high-performance models typically have a large number of parameters, making them unsuitable for running on resource-constrained devices.

[0007] To achieve lightweight models, various methods have been proposed in the industry, such as model pruning, quantization (e.g., quantizing FP32 to INT8), knowledge distillation, and designing lightweight network architectures (e.g., MobileNet, ShuffleNet). These methods aim to reduce the computational and storage overhead of the model, enabling it to run on embedded devices. However, simple lightweighting often leads to a decrease in model performance. How to maintain or improve recognition accuracy while achieving lightweighting, especially the ability to capture subtle facial features of specific target groups (such as teenagers), remains a pressing technical challenge.

[0008] In the field of object detection, the YOLO (You Only Look Once) series of models are widely popular for their high efficiency and accuracy. The latest YOLO11 model has further improved performance, but its direct application to emotion recognition tasks, which require both lightweight deployment and high accuracy on resource-constrained devices, still requires targeted improvements and optimizations.

[0009] In summary, existing technologies still have room for improvement in achieving efficient and accurate emotion recognition in adolescents while meeting the requirements for lightweight deployment on resource-constrained devices. This invention aims to provide a solution to overcome the shortcomings of existing technologies. Summary of the Invention

[0010] This invention aims to address two major problems in existing methods for recognizing emotions in adolescents: First, traditional emotion recognition models lack the ability to extract features when dealing with the complex and subtle facial expressions of adolescents, resulting in limited recognition accuracy; second, existing high-performance models typically have high computational complexity, making it difficult to achieve real-time, efficient, and lightweight deployment on resource-constrained edge devices.

[0011] To address the aforementioned technical problems, this invention proposes a method and system for adolescent emotion recognition based on an improved YOLO11 model. This invention achieves significant improvements in the accuracy and real-time performance of adolescent emotion recognition by deeply improving and optimizing the YOLO11 model, combined with specific training strategies and deployment techniques, while maintaining the model's lightweight nature.

[0012] This invention provides a method for efficient identification of adolescent emotions, comprising the following steps:

[0013] S1: Video Stream Acquisition and Sparse Keyframe Decoding. Video streams are acquired in real-time via camera modules integrated into monitoring, door locks, or similar computing devices. To balance real-time performance and computational resources, this invention designs a periodic interrupt mechanism to intelligently decode the video stream and acquire sparse keyframes. This interrupt mechanism is configured to trigger synchronously during the processor's inference idle window, ensuring image processing is performed without affecting the device's main tasks. Configurable sampling rates (e.g., 5 frames / second to 15 frames / second, or customized according to actual needs) allow the system to dynamically adjust the processing frequency based on device performance and application scenarios, thereby effectively reducing the computational burden of subsequent processing while ensuring information integrity.

[0014] S2: Device-side face detection and single-target face image extraction. After receiving sparse keyframes, face detection is performed on the device. This step can employ a lightweight face detection module pre-installed on the device (such as a lightweight version of MTCNN or RetinaFace) or a face detection algorithm specifically optimized for edge devices to automatically identify and accurately select all face regions in the frame. For each detected face in a frame, the system performs cropping and normalization processing to generate single-target face image data for subsequent emotion recognition processing. This step ensures that the emotion recognition model focuses only on face regions, reducing interference from background noise.

[0015] S3: Adolescent Emotion Recognition Based on Improved YOLO11 Model. The single-target face image data extracted in step S2 is input into the adolescent emotion recognition model (YOLO11-QX) deployed on the device. This model is the core innovation of this invention. It is built based on the YOLO11 backbone network and has made the following key improvements and optimizations for the adolescent emotion recognition task:

[0016] S31: The SOCA module is deeply integrated into the backbone network of the YOLO11 model. The SOCA module (Cross-Scale Perceptual Self-Attention) is a novel feature extraction component proposed in this invention, designed to overcome the limitations of traditional convolutional networks in capturing subtle facial expressions in adolescents. Through its unique cross-scale feature fusion mechanism, the SOCA module effectively integrates feature information from different visual fields, comprehensively capturing facial features from coarse to fine granular. Simultaneously, its built-in multi-head self-attention mechanism allows the model to dynamically focus on the relationships and changes between different expression key points (such as eyebrows, eyes, and corners of the mouth) within the facial region, thereby significantly enhancing the model's sensitivity to facial expression changes and its robustness to complex facial expressions. This design enables the YOLO11-QX model to more accurately identify the unique and subtle emotional expressions of adolescents, such as slight upturns of the mouth and subtle adjustments to the eyebrows—features that might be overlooked in traditional models.

[0017] S32: Using the DIoU Loss Function for Bounding Box Regression Optimization. During model training, this invention uses the DIoU (Distance Intersection over Union) loss function instead of the traditional IoU (Intersection over Union) or GIoU (Generalized IoU) loss functions for bounding box regression optimization. The DIoU loss function, while considering bounding box overlap, further considers the distance between the center points of the predicted box and the ground truth box and their directionality, making the regression process more stable and efficient. This helps the model achieve more accurate bounding box regression in face detection and emotion region localization, improving localization accuracy even under changes in face pose or occlusion, and accelerating the convergence speed of model training, thereby obtaining model weights with superior performance.

[0018] S33: Lightweight Model and Parallel Output To meet the lightweight deployment requirements of resource-constrained devices, the YOLO11-QX model has undergone specific optimizations. These include, but are not limited to, ONNX format conversion, INT8 quantization (converting model weights and activation values ​​from floating-point numbers to 8-bit integers), and / or model pruning. INT8 quantization significantly reduces model size and lowers computational complexity and memory bandwidth requirements during inference, while substantially improving inference speed while maintaining acceptable accuracy loss. With these optimizations, the model can perform efficient inference with extremely low latency on edge devices. The model outputs in parallel the adolescent emotion categories (e.g., anger, sadness, happiness, neutrality, surprise, fear, disgust, etc.) corresponding to the identified single-target faces, along with the prediction confidence for each emotion, providing quantitative basis for subsequent analysis and decision-making.

[0019] S4: Emotion Detection Result Encapsulation and Uploading. The emotion category, prediction confidence, and corresponding face bounding box coordinates obtained in step S3 are structurally encapsulated. To track the same target in consecutive video frames, this invention introduces a tracking identifier (user_id). This user_id is an ID used to uniquely identify a single target face appearing consecutively in the video stream. The encapsulated data fields include at least: tracking identifier (user_id), emotion category (emotion_class), confidence, and bounding box coordinates (bbox[x,y,w,h]). These single-target emotion detection results are uploaded in real time to the backend server via the device's communication interface (such as Wi-Fi, 4G / 5G, Ethernet, etc.) for unified storage, analysis, and alerting.

[0020] S5: Model Training Process (Applicable to Model Claims) For the training process of the adolescent emotion recognition model, this invention adopts the following strategy:

[0021] S51: Dataset Selection and Construction. Based on large-scale public datasets (such as AffectNet, RAF-DB, etc.), combined with 5000 self-made datasets, image samples that better match the characteristics of teenagers were selected. The existence of the self-made dataset, especially the specific data for the teenager group, effectively makes up for the lack of teenager facial expression data in general datasets, which can help the model capture more unique details and facial expression patterns, thereby improving the model's generalization ability and accuracy in actual teenager emotion recognition tasks.

[0022] S52: Refined Emotion Labeling. The image samples selected and retained in step S51 are then meticulously labeled with seven emotion categories. These seven emotion categories are expanded based on the Ekman basic emotion model, including, for example: anger, sadness, happiness, neutrality, surprise, fear, and disgust. Through professional and standardized labeling, the quality and consistency of the training dataset are ensured, providing accurate supervision information for the model to learn the semantic features of different emotions.

[0023] S53: End-to-end Model Training and Evaluation. Using the adolescent emotion image dataset annotated in step S52 as the training set, an end-to-end training process was performed on the adolescent emotion recognition model based on the improved YOLO11 backbone network (integrating the SOCA module and employing the DIoU loss function). The training process employed appropriate optimizers (such as Adam, SGD, etc.), learning rate scheduling strategies, and regularization techniques until the model reached the preset performance metrics and converged. To ensure the model's generalization ability and stability, scientific evaluation methods such as multi-fold cross-validation were used during training, and model selection was based on the validation set performance.

[0024] The adolescent emotion recognition model provided by this invention is characterized by inheriting the high efficiency of YOLO11 and, through the aforementioned structural and training optimizations, being able to receive single-target face image data as input and accurately output the adolescent emotion category and its confidence level for the recognized face. The deeply integrated SOCA module in the backbone network significantly enhances the model's accuracy in extracting subtle facial features and its robustness to complex facial expression changes; the adoption of the DIoU loss function achieves more accurate face region localization and higher training convergence efficiency; and optimizations for resource-constrained devices (such as INT8 quantization or pruning) ensure lightweight deployment of the model while maintaining high accuracy.

[0025] This invention has significant advantages over existing technologies. Through the above technical solution, this invention:

[0026] 1. Significantly improves the accuracy of emotion recognition in adolescents: The SOCA module introduced into the backbone network, combined with its unique cross-scale feature fusion and multi-head self-attention mechanism, can capture the complex and subtle facial expression features of adolescents more meticulously and robustly, effectively improving the accuracy of recognition.

[0027] 2. Achieving lightweight and efficient model deployment: Through optimization measures such as ONNX conversion and INT8 quantization, the model of this invention maintains high accuracy while significantly reducing model size and computational complexity, enabling it to be efficiently deployed on resource-constrained edge devices (such as surveillance camera modules, door lock devices, etc.) to meet real-time inference requirements.

[0028] 3. Improve bounding box localization accuracy and training efficiency: By using the DIoU loss function for bounding box regression optimization, not only is more accurate face region localization achieved, but the convergence speed of the model is also accelerated, making model training more efficient and ultimately obtaining a model with better performance.

[0029] 4. Supports large-scale, distributed applications: The combination of sparse keyframe decoding mechanism, lightweight device deployment, and centralized backend upload enables the present invention to build a distributed, lightweight psychological emotion recognition network in scenarios such as schools, communities, and families, providing reliable and real-time technical support for early psychological intervention for adolescents.

[0030] 5. Stronger generalization ability and robustness: The training strategy based on a combination of public datasets and self-made adolescent-specific datasets, as well as refined seven-category emotion annotations, enables the model to learn richer and more representative facial expression features, thereby demonstrating stronger generalization ability and robustness to changes in lighting, pose, etc. in diverse real-world application scenarios.

[0031] For example, the model trained by this invention performs excellently on key evaluation metrics, with mAP50(B) reaching 0.94557 and mAP50-95(B) reaching 0.81662. These quantitative data demonstrate that this invention significantly improves the accuracy and effectiveness of emotion recognition in adolescents while maintaining lightweight performance. Detailed Implementation

[0032] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0033] This invention discloses a method and system for adolescent emotion recognition based on an improved YOLO11 model. This method and system can efficiently and accurately identify adolescent emotions and supports lightweight deployment on resource-constrained devices. The proposed method and system mainly include modules for video stream acquisition and sparse keyframe decoding, device-side face detection and single-target face image extraction, adolescent emotion recognition based on the improved YOLO11 model, and emotion detection result encapsulation and uploading.

[0034] 1. Example 1: A Highly Efficient Method for Identifying Adolescent Emotions

[0035] The present invention provides a method for efficient identification of adolescent emotions, comprising the following steps:

[0036] S1: Acquire the video stream and decode sparse keyframes using a periodic interrupt mechanism.

[0037] The computing device used in this invention (e.g., an embedded processor and camera module built into a surveillance camera, smart lock, or tablet computer) acquires video streams in real time through its camera module. To optimize processing efficiency and resource utilization, this invention designs an intelligent sparse keyframe decoding mechanism:

[0038] Hardware Deployment: In actual deployment, computing devices are equipped with custom chip development boards based on ARM architecture, which typically integrate independent image signal processors (ISPs) and hardware codecs, as well as NPU units for accelerating AI inference.

[0039] Periodic interrupts: Video stream decoding is not performed continuously, but rather through a periodic interrupt mechanism to achieve sparse sampling. This interrupt mechanism is programmed to synchronize with the processor's inference "idle window." For example, an interrupt is triggered after the processor completes the current emotion recognition inference task, or when the system load is low, thus instructing the hardware decoder to extract a frame from the video stream.

[0040] Configurable Sampling Rate: The sampling rate can be dynamically configured according to actual application needs, device performance, and real-time requirements. For example, in scenarios with high real-time requirements, the sampling rate can be configured to 15 frames per second; in scenarios sensitive to privacy protection or resource consumption, it can be configured to 5 frames per second. In this embodiment, the system defaults to a sampling rate of 10 frames per second to balance the ability to capture emotional changes with computational resource consumption. This sparse sampling significantly reduces the amount of image data that needs to be processed, thereby reducing the computational pressure on subsequent face detection and emotion recognition. The acquired frames are typically in RGB or grayscale image format, and the resolution can be determined according to the camera specifications and processing capabilities, such as 1080P or 720P.

[0041] S2: The device performs face detection and extracts single-target face image data.

[0042] After acquiring the sparse keyframe, the computing device immediately performs face detection on the frame to quickly locate and extract all face regions in the frame.

[0043] Face detection module: This embodiment can use a pre-trained lightweight face detection model, such as a lightweight model based on SSD (Single Shot MultiBox Detector) after pruning and quantization. These models are pre-deployed in the device.

[0044] Real-time detection and cropping: For each captured image frame, the face detection module identifies the faces appearing in the scene in real time and outputs the bounding box coordinates ([x1, y1, x2, y2]) and confidence score of each detected face.

[0045] Generating single-target face images: The system crops the original frame image based on the detected face bounding boxes to extract a single target face region. To meet the input requirements of the subsequent emotion recognition model, these cropped face images are typically scaled to a uniform size (e.g., 224x224 or 112x112 pixels) and normalized. If multiple faces are detected in a frame, subsequent emotion recognition is performed on each face region separately.

[0046] S3: An emotion recognition model for teenagers deployed on the device performs emotion recognition.

[0047] The single-target face image data generated in step S2 is input in real time to the YOLO11-QX adolescent emotion recognition model deployed on a computing device. This model is the core of this invention, and its detailed structure and characteristics are as follows:

[0048] S31: The YOLO11-QX model is based on the backbone network of YOLO11, but it deeply integrates the SOCA (Cross-Scale Perceptual Self-Attention) module proposed in this invention into its intermediate or bottleneck layers.

[0049] SOCA module structure: This module typically consists of the following parts:

[0050] Multi-scale feature extractor: Contains several parallel convolutional branches and / or pooling layers to extract information from the input feature map from different receptive fields (i.e., different scales). For example, it may contain 1x1, 3x3, and 5x5 convolutional branches, as well as an average pooling branch.

[0051] Feature fusion unit: Through operations such as interpolation, concatenation and / or element-wise addition, feature information at different scales is effectively fused to generate a unified multi-scale feature representation.

[0052] Multi-head self-attention mechanism: A multi-head self-attention mechanism is applied to the fused feature map. This mechanism divides the fused feature map into multiple "heads," each of which independently computes the query, key, and value. It models the dependencies between feature channels or spatial locations through scaled dot-product attention. This allows the model to dynamically emphasize locally sensitive regions (such as eyebrows, corners of the eyes, and corners of the mouth) that are crucial for adolescent emotion recognition, while suppressing irrelevant background information.

[0053] Integration Method: The SOCA module is embedded into different stages of the YOLO11 backbone network, such as in residual blocks before downsampling or in specific layers of the Feature Pyramid Network (FPN) structure. This ensures that the model can benefit from both high-level semantics and low-level details simultaneously, thereby significantly enhancing the model's accuracy in extracting facial features and its robustness to complex facial expression changes. For example, in the YOLO11 CSPDarknet backbone network, the SOCA module can be inserted after the last bottleneck layer in the second, third, and fourth stages.

[0054] S32: Using the DIoU loss function for bounding box regression optimization. During the training of the YOLO11-QX model, the bounding box regression loss of its object detection branch uses the DIoU (Distance Intersection over Union) loss function.

[0055] DIoU Calculation: The DIoU loss function considers not only the IoU overlap between the predicted and ground truth boxes, but also the Euclidean distance between the center points of the two boxes, accelerating the regression process by minimizing the center point distance. Its formula is as follows: ,in It is intersection, union, and comparison. It is the center point of the prediction box. Center point of the real frame The square of the Euclidean distance between them It is the diagonal length of the smallest enclosed rectangle that simultaneously contains the predicted bounding box and the ground truth bounding box.

[0056] Advantages: Compared to traditional IoU or GIoU, the DIoU loss function provides a more stable gradient during the regression process. Especially when the predicted box and the ground truth box do not overlap, it can still provide an effective optimization direction, thereby accelerating the convergence efficiency of model training. This allows the model to more accurately locate the facial region of teenagers, and maintain high accuracy even when facial expressions and movements cause deformation of the face boundary.

[0057] S33: Lightweight Model Processing and Parallel Output To ensure that the YOLO11-QX model can run efficiently on resource-constrained devices, this invention has implemented the following lightweight optimizations:

[0058] ONNX Format Conversion: After training, the .pt format model is first converted to the ONNX (Open Neural Network Exchange) standard format. ONNX is an open deep learning model representation format that facilitates model deployment across different frameworks and hardware platforms.

[0059] INT8 Quantization: The ONNX format model is further quantized using INT8 (8-bit integer) quantization. Quantization is a key technology in this invention that significantly improves inference speed and reduces power consumption. By converting the model's floating-point weights and activation values ​​to 8-bit integers, the model size and runtime memory bandwidth are reduced, and it allows the use of computationally optimized AI accelerators (such as Neural Processing Units, NPUs) in modern processors. The quantization process can employ various strategies, such as Post-Training Quantization (PTQ) or Quantization-Aware Training (QAT). This embodiment prioritizes PTQ, using a calibration dataset to perform statistical analysis on each layer of the model to determine appropriate quantization parameters (scaling factor and zeros) to minimize accuracy loss.

[0060] Parallel Output: The lightweight optimized YOLO11-QX model can output the emotion category and its prediction confidence for each single target face in parallel. Emotion categories include, for example, seven: anger, sadness, happiness, neutrality, surprise, fear, and disgust. The confidence score is a floating-point number between 0 and 1, representing the model's certainty in predicting that emotion.

[0061] S4: Packaging and Uploading of Emotion Detection Results

[0062] After step S3 completes emotion recognition, the system will encapsulate the recognition results in a structured manner and upload them to the backend server through the device's communication interface.

[0063] Data encapsulation: To ensure data integrity and traceability, the encapsulated data must include at least the following fields:

[0064] Tracking identifier (user_id): Used to uniquely identify the same teenager in consecutive frames. The user_id can be generated based on face tracking algorithms (such as SORT, DeepSORT, etc.) to identify and associate the same face within a time series. This allows the backend server to perform long-term monitoring and trend analysis of the emotions of specific teenagers.

[0065] Emotion Class (emotion_class): The identified discrete emotion class.

[0066] Confidence: The probability value that the model predicts for this emotion category, for example, 0.92.

[0067] Bounding box coordinates (bbox[x,y,w,h]): The position and size information of the face in the original frame for identifying emotions.

[0068] Timestamp: The specific time when the frame was captured and identified, facilitating data analysis.

[0069] Communication Interface and Upload: The encapsulated data packets are transmitted encrypted through the device's communication interface (e.g., but not limited to Wi-Fi, 4G / 5G cellular networks, Ethernet, Bluetooth, etc.). To ensure data security and interface authentication, an interface signature mechanism based on the Chinese national cryptographic algorithm SM3 is used to sign the data packets, while the data content is encrypted using the AES (Advanced Encryption Standard) algorithm. The encrypted data packets are then securely transmitted to the backend server. The backend server can be a cloud server, primarily used to receive, store, and further analyze the adolescent emotional data from edge devices.

[0070] 2. Example 2: Adolescent Emotion Recognition Model

[0071] This invention provides a YOLO11-QX model for adolescent emotion recognition. It is a deep learning model built upon the YOLO11 backbone network and specifically trained and optimized for adolescent emotion recognition tasks. The model is configured to receive preprocessed single-target face image data as input and output the adolescent emotion category of the recognized face and its confidence level.

[0072] Model Architecture: The core of YOLO11-QX is an improved YOLO11 network, characterized by:

[0073] Backbone network integration of SOCA module: As described in Example 1, S31, a custom SOCA module is embedded in the YOLO11 backbone network, for example, after the C3 module or in a specific feature extraction stage. This module enhances the model's ability to extract fine features from key facial regions through parallel multi-scale convolution and multi-head self-attention mechanisms, exhibiting stronger robustness, especially when dealing with subtle facial expression changes in teenagers.

[0074] DIoU loss function optimization: The model's loss function includes DIoU loss as a component of bounding box regression, ensuring that the model can learn the localization of face regions more accurately during training.

[0075] Model training: Model training follows these steps:

[0076] S51: Dataset Selection and Construction: The training data is based on large-scale public sentiment datasets (such as AffectNet, FER2013, RAF-DB, SFEW, etc.), and is enhanced by combining 20,000 open datasets unique to this invention (which may be obtained from the Internet or partner institutions, but have undergone rigorous screening and adolescent age matching) and 5,000 self-made datasets. The self-made dataset consists of images collected for specific adolescent groups and scenarios, which helps to capture the emotional expression characteristics of specific age groups and cultural backgrounds, significantly improving the model's generalization ability and accuracy. Data filtering ensures that all image samples are consistent with the characteristics of adolescent age groups.

[0077] S52: Fine-grained annotation of seven emotion categories: The selected image samples are fine-grained and classified into seven emotion categories: anger, sadness, happiness, neutrality, surprise, fear, and disgust. The annotation process follows the Ekman basic emotion model for extended classification, ensuring consistency and accuracy of annotations and providing high-quality ground truth labels for supervised learning.

[0078] S53: End-to-End Model Training: The labeled dataset of adolescent emotion images was used as the training set, and the YOLO11-QX model was trained end-to-end on a high-performance computing server (e.g., equipped with dual NVIDIA Tesla T4 GPUs, 32GB of memory, and an Intel Xeon processor). The Adam optimizer was used during training, with an initial learning rate set to 0.001 and a cosine annealing learning rate scheduling strategy. The batch size was set to 100. Data augmentation techniques, such as random flipping, random cropping, and color dithering, were also employed during training to improve the model's generalization ability. Training continued until the model performance stabilized on the validation set and met the preset performance metrics: MPA@0.50 reached 0.94557, and mAP@0.95 reached 0.81662.

[0079] Model Deployment Optimization: As described in Example 1, S33, the trained model undergoes ONNX conversion and INT8 quantization. This optimization significantly reduces the model size, for example, from hundreds of MB to tens of MB in INT8, and increases inference speed several times, enabling efficient deployment on resource-constrained devices, such as NVIDIA Jetson series, Raspberry Pi, and other embedded platforms.

[0080] 3. Example 3: Computing Device and Computer-Readable Storage Medium

[0081] The present invention also provides a computing device capable of implementing the above-mentioned efficient method for recognizing adolescent emotions, and a computer-readable storage medium storing the program of the method.

[0082] Computing devices: such as Figure 3 As shown in the diagram (assuming a structural schematic of a computing device), this computing device can be a smart camera, a smart door lock, a smart learning machine, or any terminal device with embedded computing capabilities. Its main components include:

[0083] Camera module: used to perform the video stream acquisition step (S1) as described in claim 1.

[0084] The processor, such as an Intel Xeon processor (for training) or an edge AI chip (for deployment), is configured to perform the face detection step (S2) as described in claim 1, the adolescent emotion recognition step (S3) by incorporating the adolescent emotion recognition model as described in claim 2, and the result encapsulation step (S4). The processor has powerful computing capabilities.

[0085] Memory: For example, 32GB of memory is used to store model weights, intermediate feature maps, and processed data.

[0086] Hard disk / storage media: Used to store the operating system, applications, and trained YOLO11-QX models.

[0087] Communication interface: such as Wi-Fi module, 4G / 5G module or Ethernet interface, configured to upload the single-target emotion detection results generated in S4 to the backend server.

[0088] Computer-readable storage medium: Specifically, the present invention provides a computer-readable storage medium having computer program instructions stored thereon. When the instructions are executed by a processor of a computing device, the processor is able to implement all steps (S1 to S4) of the efficient adolescent emotion recognition method of claim 1. The storage medium may be a read-only memory (ROM), a hard disk, or a similar non-transitory storage medium.

[0089] Through the detailed description of the above specific embodiments, it is clear that the core innovations, technical details, and potential application value of this invention in the field of adolescent emotion recognition can be clearly seen. Attached image description: Figure 1 Flowchart of methods for detecting adolescent emotions Figure 2 Schematic diagram of the YOLO11-QX model network structure Figure 3 Illustration of adolescent emotion labels Figure 4 Label correlation plot of YOLO11-QX model training results Figure 5 Metrics chart of YOLO11-QX model training results Figure 6 Image showing the recognition results of the YOLO11-QX model.

Claims

1. A method for efficiently identifying emotions in adolescents, characterized in that, Includes the following steps: S1: Acquire a video stream and decode the video stream using a periodic interrupt mechanism designed to synchronize with the processor's inference idle window to obtain sparse keyframes. The periodic interrupt mechanism can be configured to a sampling rate of 5 frames / second to 15 frames / second or a custom sampling rate. S2: On the device, face detection is performed on the sparse keyframes to automatically identify and extract the detected face regions, thereby generating single-target face image data for subsequent processing; S3: On the device, the single-target face image data is input into a youth emotion recognition model deployed therein. The model is built on the YOLO11 backbone network and has the following improved features: S31: The backbone network of the model is deeply integrated with the SOCA module. The SOCA module aims to significantly enhance the model’s accuracy in extracting facial features and its robustness to complex facial expression changes through its unique cross-scale feature fusion and multi-head self-attention mechanism. S32: The model uses the DIoU loss function to optimize bounding box regression during training to achieve more accurate face region localization and improve the convergence efficiency of model training. S33: The model has undergone specific optimizations (e.g., INT8 quantization or pruning) to meet the requirements of lightweight deployment on resource-constrained devices, and outputs in parallel the category of adolescent emotions corresponding to the single target face and its prediction confidence. S4: Encapsulate the emotion category, confidence level, and corresponding face bounding box coordinates, and include them in a tracking identifier (user_id) in consecutive frames to uniquely identify the single target face, forming a single-target emotion detection result, and upload the result to the backend server.

2. A model for recognizing emotions in adolescents, characterized in that: The model is built on the YOLO11 backbone network and has been specifically trained and optimized for adolescent emotion recognition tasks. The model is configured to receive single-target face image data as input and output the adolescent emotion category of the recognized face and its confidence level. The model includes the following structure and / or features: The model backbone network is deeply integrated with the SOCA module, which aims to significantly enhance the model's accuracy in extracting subtle facial features and its robustness to complex facial expression changes through its unique cross-scale feature fusion and multi-head self-attention mechanism. The model achieves more accurate face region localization and exhibits higher training convergence efficiency by using the DIoU loss function for bounding box regression optimization during training. The model has been optimized for resource-constrained devices (e.g., INT8 quantization and / or pruning) to achieve lightweight deployment while maintaining high accuracy.

3. The method or model according to claim 1 or 2, wherein, The training process of the model includes the following steps: S51: Based on public and private datasets, filter and retain image samples that match the characteristics of the adolescent age group; S52: Perform emotion labeling on the preserved image samples. The labeled emotion categories can be configured and adjusted according to actual application needs or subsequent research results. For example, it includes seven emotions: anger, sadness, happiness, neutrality, surprise, fear, and disgust, to form a refined training dataset specifically for the emotional characteristics of adolescents. S53: Using the labeled adolescent image dataset as the training set, perform end-to-end training on the model until the model reaches the preset performance index and achieves convergence.

4. The method or model according to claim 1 or 3, wherein, The face detection step is implemented using a pre-installed face detection module on the device or other face detection modules to identify and extract face regions in the image.

5. The method or model according to claim 1 or 3, wherein, The emotion labeling mentioned in S52 follows the Ekman basic emotion model for extended classification; and the model performance is scientifically evaluated using methods such as multi-fold cross-validation.

6. The method according to claim 1, wherein, The encapsulated data fields mentioned in S4 include at least the following: a tracking identifier (user_id), emotion category (emotion_clss), confidence level, and bounding box coordinates (bbox[x,y,w,h]) used to uniquely identify the single target face in consecutive frames.

7. A computing device, characterized in that, include: A camera module configured to perform the video stream acquisition step (S1) as described in claim 1. The processor, memory, and hard disk are configured to perform the face detection step (S2) as described in claim 1, the adolescent emotion recognition step (S3) by carrying the adolescent emotion recognition model as described in claim 2, and the execution result encapsulation step (S4). The communication interface is configured to upload the single-target emotion detection results generated in S4 to the backend.

8. A computer-readable storage medium, characterized in that, It stores computer program instructions, which, when executed by the processor of the computing device, cause the processor to implement all steps (S1 to S4) of the efficient adolescent emotion recognition method according to claim 1.

Images