AR auxiliary earphone configuration method for charging box screen

Through adaptive multi-scale attention U-Net model and reinforcement learning technology, combined with AR interaction, the accuracy and personalization problems of the existing charging box in the headphone configuration are solved, and an efficient and personalized headphone configuration experience is achieved, which improves the User satisfaction and product competitiveness.

CN119068146APending Publication Date: 2024-12-03COSONIC INTELLIGENT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410918680.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-10
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

During the headset configuration process, the existing charging box with screen has problems such as low AR guidance accuracy, insufficient personalization, poor real-time performance, lack of adaptability and single interaction methods, resulting in poor user experience.

Method used

Adaptive multi-scale attention U-Net model combined with reinforcement learning, through deep learning technology and AR interaction, the ear and headphone image processing is performed to generate personalized AR guidance information, and optimize configuration strategies through user feedback to support multimodal interaction Such as touch, voice and gesture recognition.

Benefits of technology

It realizes high-precision and personalized headphone configuration guidance, improves user experience, provides smooth real-time AR interaction, quickly adapts to new models of headphones and user habits, continuously optimizes configuration strategies, significantly improving user satisfaction and product competitiveness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068146A_ABST
    Figure CN119068146A_ABST
Patent Text Reader

Abstract

The invention relates to an AR auxiliary earphone configuration method for a charging box screen. The method comprises the following steps: obtaining images of ears and earphones; preprocessing the image; processing the preprocessed image by using a self-adaptive multi-scale attention U-Net model; carrying out feature analysis and matching based on the output of the AMA-UNet model; generating AR guidance information according to feature analysis and matching results; displaying AR guidance information on a screen of the charging box, and receiving feedback of a user on earphone position adjustment; based on user feedback, optimizing a configuration strategy by using a reinforcement learning algorithm and updating the AMA-UNet model; and verifying an optimized configuration result, if the configuration result does not reach a preset threshold value, returning to the step of generating the AR guidance information to carry out iterative optimization until the preset threshold value is reached or a preset number of iterations is completed, and by adopting an innovative AMA-UNet model, the AR configuration guidance precision is improved to a millimeter level, and the configuration effect is greatly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of AR-assisted earphone configuration methods, and more specifically, to an AR-assisted earphone configuration method for a charging box screen. Background Art

[0002] With the rapid development of smart wearable devices, wireless headphones have become an indispensable electronic product in daily life. In recent years, in order to enhance user experience and functional diversity, a new type of headphone charging box with screen has gradually entered the market, ushering in a new era of headphone accessories. This innovative design not only provides users with a more intuitive information display, but also opens up new possibilities for intelligent management and personalized settings of headphones.

[0003] The current status of the existing technology is as follows:

[0004] 1. Basic information display charging box: The original charging box with a screen mainly displays basic information such as battery level and connection status. Although this design provides more information than the traditional screenless charging box, its functions are still very limited.

[0005] 2. Touch-operated charging box: The subsequent charging box with a touch screen allows users to perform simple operations such as adjusting the volume, switching music, etc. This increases the interactivity of the charging box, but still does not fully utilize the potential of the screen.

[0006] 3. Health monitoring integrated into charging box: Some high-end products have begun to integrate health monitoring functions on the charging box screen, such as displaying heart rate, step counting data, etc. This expands the functional scope of the charging box, but the connection with the core experience of the headset is still insufficient.

[0007] 4. Simple AR guidance charging box: Some of the latest products try to provide simple AR guidance on the charging box screen to assist with pairing and basic settings of earphones. This marks the beginning of the use of the charging box screen to improve the earphone experience.

[0008] The existing technology has the following problems: Taking a simple AR guidance charging box as an example, although it represents the forefront of current technology, it still has the following obvious problems:

[0009] 1. Low AR guidance accuracy: Existing systems usually use basic image processing algorithms, which cannot accurately identify and analyze the complex structure of the ear and earphones, resulting in inaccurate AR guidance.

[0010] 2. Insufficient personalization: Most systems use a universal guidance model that cannot adapt to the differences in ear shapes and personal preferences of different users, reducing the practicality of the guidance.

[0011] 3. Poor real-time performance: Due to the limited processing power of the charging box, complex AR algorithms are difficult to run in real time, resulting in obvious delays and freezes, affecting the user experience.

[0012] 4. Lack of adaptive capabilities: The existing system is difficult to quickly adapt to new models of headphones or changes in user habits, and each update requires a lot of redevelopment.

[0013] 5. Single interaction method: Most systems only support touch operation, which is inconvenient during the headset configuration process, especially for users who need to adjust the headset at the same time.

[0014] 6. Lack of continuous optimization mechanism: Most existing systems are static in design and lack the ability to continuously learn and optimize based on user feedback, and are unable to provide increasingly accurate personalized experiences.

[0015] Considering the limitations of existing charging box technology with screens, there is an urgent need for a headset configuration solution that can make full use of the charging box screen and provide high-precision, personalized, real-time and easy-to-use. Summary of the invention

[0016] In view of the above problems existing in the prior art, the AR-assisted headset configuration method of the charging box screen proposed in the present invention is designed to solve these problems. By combining advanced deep learning technology, AR interaction and reinforcement learning optimization, the present invention aims to give full play to the potential of the charging box screen and completely change the headset configuration and usage experience.

[0017] The present invention provides an AR-assisted headset configuration method for a charging box screen, comprising the following steps:

[0018] Obtain images of ears and headphones;

[0019] Preprocessing the image;

[0020] Processing the preprocessed image using an adaptive multi-scale attention U-Net model;

[0021] Perform feature analysis and matching based on the output of the AMA-UNet model;

[0022] Generate AR guidance information according to the result of the feature analysis and matching;

[0023] Displaying the AR guidance information on the charging box screen and receiving user feedback on the adjustment of the earphone position;

[0024] Based on the user feedback, optimizing the configuration strategy and updating the AMA-UNet model using a reinforcement learning algorithm;

[0025] The optimized configuration result is verified. If the preset threshold is not reached, the step of generating AR guidance information is returned to perform iterative optimization until the preset threshold is reached or the predetermined number of iterations is completed.

[0026] Specifically, the step of obtaining the image of the ear and the earphone comprises:

[0027] Use image sensors to capture images at 4K resolution;

[0028] The exposure parameters are adjusted using an adaptive exposure control algorithm, wherein the adaptive exposure control algorithm is defined as:

[0029] E=f(L,S,N)

[0030] f(L,S,N)=α·log(L)+β·S-γ·N 2

[0031] Among them, E is the exposure value, L is the ambient light intensity, S is the sensor sensitivity, N is the noise level, and α, β, and γ are adaptive parameters.

[0032] Specifically, the step of preprocessing the image adopts a multi-scale adaptive contrast enhancement algorithm, which is defined as:

[0033] I out =G(I in ,s,r)

[0034] G(I,s,r)=I+r·(IF(I,s))

[0035] Among them, I in and I out are the input and output images respectively, s is the scale factor, r is the enhancement strength, and F(I,s) is the Gaussian blur function.

[0036] Specifically, the AMA-UNet model includes:

[0037] Adaptive multi-scale feature extraction module; dynamic attention mechanism; lightweight depth-separable convolution and adaptive knowledge distillation module.

[0038] Specifically, the adaptive multi-scale feature extraction module adopts a dynamic convolution kernel generation network, which is defined as:

[0039] K i =DKGN(X,i)

[0040] F ms =∑(w i Conv(X,K i )),i∈{1,3,5,7}

[0041] Among them, DKGN(X,i) dynamically generates the convolution kernel according to the input X and the target kernel size i, w i Specifically, the dynamic attention mechanism introduces the spatiotemporal attention mechanism, which is defined as:

[0042] A=σ(f c (X)·f s (X)·f t (X)

[0043] F att =A·X

[0044] Among them, f c ,f s ,f t are the channel, spatial and temporal attention functions respectively, and σ is the sigmoid function. Specifically, the lightweight depth-wise separable convolution introduces an adaptive activation function, which is defined as:

[0045] F dw =AAF(DWConv(X))

[0046] F pw =AAF(PWConv(F dw ))

[0047] AAF(x)=α·x·sigmoid(β·x)+γ·x

[0048] Among them, DWConv and PWConv are depth convolution and point-by-point convolution respectively, and α, β, and γ are learnable parameters.

[0049] Specifically, the adaptive knowledge distillation module adopts adversarial knowledge distillation, which is defined as:

[0050] L kd =α·L ce +(1-α)·D KL (S(x),T(x))+λ·L adv (S(x),T(x))

[0051] L adv =E[log(D(T(x)))]+E[log(1-D(S(x)))]

[0052] Among them, L ce is the cross entropy loss, D KL is the KL divergence, S(x) and T(x) are the outputs of the student model and the teacher model respectively, D is the discriminator, and λ is the weight factor.

[0053] Specifically, the feature analysis and matching step adopts an improved multimodal feature fusion algorithm, which is defined as:

[0054] F=Φ(F visual ,F geometric ,F semantic )Φ(V,G,S)=σ(W v ·V+W q ·G+W s ·S)

[0055] Among them, F v isual,F geometric ,F semantic are visual, geometric and semantic features respectively,

[0056] W v ,W g ,W s is a learnable weight matrix; the feature matching adopts an improved RANSAC algorithm

[0057] Method, introduce adaptive threshold:

[0058] θ=μ·exp(-λ·t)

[0059] Among them, θ is the matching threshold, t is the number of iterations, and μ and λ are adjustable parameters.

[0060] Specifically, the step of generating AR guidance information adopts a dynamic vector field generation algorithm, which is defined as:

[0061]

[0062] Among them, V(x,y) is a 2D vector field, Ψ(x,y) is a scalar potential field, Φ(x,y) is a vector potential field, and curl represent the gradient and curl operators respectively; the optimization configuration strategy adopts the meta-reinforcement learning algorithm and is defined as:

[0063] Among them, θ is the model parameter, α, β are the learning rates, L is the loss function, and f θ A multi-objective evaluation function is introduced for the step of verifying the configuration result described in the parameterized policy function, which is defined as:

[0064] S=w1·C+w2·F+w3·U

[0065] Among them, S is the overall score, C is the fit score, F is the functionality score, U is the user comfort score, and w1, w2, and w3 are weight factors.

[0066] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0067] 1. High-precision AR guidance: Using the innovative AMA-UNet model, the AR configuration guidance accuracy is improved to the millimeter level, greatly improving the configuration effect.

[0068] 2. Personalized experience: Through adaptive multi-scale feature extraction and dynamic attention mechanism, the system can accurately adapt to the ear shape characteristics and usage habits of each user.

[0069] 3. Real-time AR on the charging box: The lightweight design enables complex AR algorithms to run in real time on the limited resources of the charging box, providing a smooth visual guidance experience.

[0070] 4. Strong adaptability: Adversarial knowledge distillation technology enables the system to quickly adapt to new models of headphones, greatly reducing the adaptation time and cost when updating products.

[0071] 5. Multimodal interaction: Combining touch, voice and gesture recognition to provide users with a more natural and convenient way of operation, especially during the headset configuration process.

[0072] 6. Intelligent continuous optimization: Through reinforcement learning algorithms, the system can continuously optimize configuration strategies based on user feedback to provide an increasingly personalized experience.

[0073] 7. Improve user satisfaction: Combining the above advantages, the present invention significantly improves the user's configuration experience and earphone usage satisfaction, and reduces discomfort and configuration time.

[0074] 8. Product differentiation provides headphone manufacturers with a significant competitive advantage, making their products stand out in the market.

[0075] 9. Expanding the functions of the charging box: This invention opens up new application directions for the charging box screen and lays the foundation for integrating more intelligent functions in the future.

[0076] In summary, the present invention not only solves the key problems of existing charging boxes with screens in terms of earphone configuration, but also fully exploits the potential of the charging box screen through innovative technology combinations. It represents an important milestone in the evolution of smart wearable device configuration technology and charging box functions, and has the potential to redefine the way users interact with wireless earphones, bringing revolutionary impacts to the entire industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of an AR-assisted headset configuration method for a charging box screen of the present invention.

[0078] Figure 2 This is a schematic diagram of the adaptive multi-scale attention U-Net model framework of the present invention. DETAILED DESCRIPTION

[0079] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0080] Unless defined otherwise, 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 invention belongs.

[0081] The present invention provides an AR-assisted earphone configuration method on the charging box screen, aiming to solve the problems of poor user experience and low configuration accuracy in the traditional earphone configuration process. The method uses augmented reality (AR) technology combined with advanced deep learning algorithms to provide users with intuitive and accurate configuration guidance on the screen of the earphone charging box.

[0082] like Figure 1-2 As shown, the AR auxiliary headset configuration method of the charging box screen of the present invention includes the following steps:

[0083] 1. Get images of ears and headphones

[0084] In this step, we use a high-resolution camera integrated on the earphone charging box to capture images of the user's ears and earphones. Specifically, we use the Sony IMX7661 / 1.5650MP image sensor with an F / 1.8 aperture lens to capture images with a 4K (3840×2160) resolution. In order to adapt to different lighting conditions, we introduce an adaptive exposure control algorithm:

[0085] E=f(L,S,N)

[0086] f(L,S,N)=αlog(L)+βS-γN 2

[0087] Where E is the exposure value, L is the ambient light intensity (unit: Iux), S is the sensor sensitivity (ISO value), and N is the noise level (normalized value between 0-1). α, β, and γ are adaptive parameters, and their value ranges are usually: α∈[0.1,1.0], β∈[0.001,0.1], γ∈[0.1,2.0];

[0088] These parameters are adjusted in real time through machine learning algorithms to adapt to different environmental conditions. For example, in low-light environments, the α value will increase relatively to improve sensitivity to lighting changes; in high-noise environments, the γ value will increase to enhance noise suppression.

[0089] 2. Preprocess the image

[0090] After acquiring the image, we need to preprocess it to improve the effect of subsequent processing. Here we use the Multi-Scale Adaptive Contrast Enhancement (MS-ACE) algorithm:

[0091] I out =G(I in ,s,r)

[0092] G(I,s,r)=I+r·(IF(I,s))

[0093] Among them, I in and I out are the input and output images, s is the scale factor, and r is the enhancement strength. F(I,s) is a Gaussian blur function, whose standard deviation σ is proportional to s, usually σ=k·s, where k is a constant with a typical value of 0.5.

[0094] The value range of s is usually [0.5, 2.0]. Smaller s values ​​can enhance local details, while larger s values ​​can enhance global contrast. The value range of r is usually [0.1, 1.0]. Larger r values ​​will produce stronger enhancement effects.

[0095] In practical applications, we can use multiple scales (for example, s1 = 0.5, s2 = 1.0, s3 = 2.0) for processing, and then fuse the results to obtain a multi-scale enhancement effect.

[0096] 3. Use the adaptive multi-scale attention U-Net (AMA-UNet) model to process the preprocessed image. The AMA-UNet model is the core of the present invention, which includes four innovative modules: adaptive multi-scale feature extraction module 1, dynamic attention mechanism 2, lightweight depth-separable convolution 3 and adaptive knowledge distillation module 4.

[0097] 3.1 Adaptive multi-scale feature extraction module 1, this module uses the dynamic convolution kernel generation network (DKGN), which is defined as follows:

[0098] K i =DKGN(X,i)

[0099] F ms =∑(w i Conv(X,K i )),i∈{1,3,5,7}

[0100] DKGN(X,i) is a small neural network that dynamically generates convolution kernels based on the input X and the target kernel size i. iis the weight of each scale, learned through the attention mechanism. This method allows the model to adaptively adjust the convolution kernel according to the input, thereby capturing features of different scales. For example, for the detailed features of the ear, a smaller convolution kernel (such as 1x1 or 3x3) may be generated, while for the overall shape of the earphone, a larger convolution kernel (such as 5x5 or 7x7) may be generated.

[0101] 3.2 Dynamic Attention Mechanism 2, we introduce the temporal and spatial attention (TSA) mechanism to enhance the model’s perception of key areas

[0102] A=σ(f c (X)·f s (X)·f t (X)

[0103] F att =A·X

[0104] Among them, f c 、f s 、f t are channel, spatial and temporal attention functions respectively, and σ is the sigmoid function. The combination of these three attention mechanisms allows the model to focus on important information in different dimensions.

[0105] For example, channel attention may focus more on the feature channels representing the contour of the earphone, spatial attention may focus on the area where the ear and the earphone touch, and temporal attention can capture the dynamic process of the user adjusting the position of the earphone.

[0106] 3.3 Lightweight Depthwise Separable Convolution 3. In order to achieve real-time processing on resource-constrained devices such as charging boxes, we introduced lightweight depthwise separable convolution and combined it with adaptive activation function (AAF):

[0107] F dw =AAF(DWConv(X))

[0108] F pw =AAF(PWConv(F dw ))

[0109] AAF(x)=α·x·σ(β·x)+γ·x

[0110] Among them, DWConv and PWConv are depthwise convolution and pointwise convolution respectively, and α, β, and γ are learnable parameters. AAF allows the model to adaptively adjust the shape of the activation function according to the input, improving the expressiveness and flexibility of the model.

[0111] Typical parameter ranges are:

[0112] α∈[0.1,2.0]

[0113] β∈[0.5,5.0]

[0114] γ∈[-0.5,0.5]

[0115] 3.4 Adaptive Knowledge Distillation Module 4. In order to enable the model to quickly adapt to new models of headphones, we introduced the Adversarial Knowledge Distillation (AKD) technology:

[0116] L kd =α·L ce +(1-α)·D KL (S(x),T(x))+λ·L adv (S(x),T(x))

[0117] L adv =E[log(D(T(x)))]+E[log(1-D(S(x)))]

[0118] Among them, L ce is the cross entropy loss, D KL is the KL divergence, S(x) and T(x) are the outputs of the student model and the teacher model respectively, D is the discriminator, and λ is the weight factor. α is usually in the range of [0.3, 0.7] and λ is in the range of [0.01, 0.1].

[0119] This approach allows the model to learn general knowledge from a pre-trained “teacher” model while remaining adaptable to new data through adversarial training.

[0120] 4. Perform feature analysis and matching based on the output of the AMA-UNet model. In this step, we use the improved multimodal feature fusion algorithm (IMM-FF):

[0121] F=Φ(F visual ,F geometric ,F semantic )

[0122] φ(V,G,S)=σ(W v ·V+W g ·G+W s ·S)

[0123] Among them, F visual 、F geometric 、F semantic are visual, geometric and semantic features respectively, and W v , W g , W s is a learnable weight matrix. This approach allows us to comprehensively consider the appearance, shape, and functional characteristics of the headphones.

[0124] For feature matching, we use the improved RANSAC algorithm and introduce an adaptive threshold:

[0125] θ=μ·exp(-λ·t)

[0126] Among them, θ is the matching threshold, t is the number of iterations, and μ and λ are adjustable parameters. μ is usually in the range of [0.5, 2.0], and λ is in the range of [0.01, 0.1]. This adaptive threshold strategy can gradually tighten the matching conditions during the iteration process and improve the matching accuracy.

[0127] 5. Generate AR guidance information based on the results of feature analysis and matching. In order to generate smooth and intuitive AR guidance information, we use the dynamic vector field generation algorithm (DVFG):

[0128]

[0129] Among them, V(x,y) is a 2D vector field, Ψ(x,y) is a scalar potential field, Φ(x,y) is a vector potential field, and curl represent the gradient and curl operators, respectively. Ψ and Φ are generated through a deep neural network to ensure that the generated AR guidance is smooth and continuous. This approach can create natural and smooth visual guidance, such as arrows indicating the direction of headset movement or rotation indicators for adjusting angles.

[0130] 6. Display AR guidance information on the charging box screen and receive user feedback on the adjustment of the earphone position. The charging box screen uses a high-resolution OLED display with a resolution of 1080x1920 to ensure the clarity of AR guidance information. User feedback is collected through multimodal methods such as touch screen, voice commands or gesture recognition. For voice command recognition, we use an improved temporal convolutional network (I-TCN):

[0131] O=I-TCN(X)

[0132] I-TCN contains dilated convolutions and residual connections, which improves the timing modeling capability and is suitable for processing voice commands.

[0133] For gesture recognition, we use a hybrid model of 3D convolutional neural network and long short-term memory network (3D-CNN-LSTM) G = 3D-CNN-LSTM (F), where F is a 3D tensor composed of consecutive frames. This method can effectively capture the spatiotemporal characteristics of gestures.

[0134] 7. Based on user feedback, we use the reinforcement learning algorithm to optimize the configuration strategy and update the AMA-UNet model. We use the meta-reinforcement learning algorithm to optimize the configuration strategy:

[0135]

[0136]

[0137] Among them, θ is the model parameter, α and β are learning rates, L is the loss function, and f θ is a parameterized policy function. α and β are usually in the range of [0.0000.01]. This approach allows the model to quickly adapt to new headphone models and user preferences, and learn effective configuration strategies with a small number of samples.

[0138] The system records the user's actions and results of each adjustment of the headphone position, including the headphone position before and after the adjustment, user satisfaction score, etc. The cumulative return for each time step is calculated using a discount factor γ (usually set to 0.9-0.99) to balance immediate and long-term benefits. The policy gradient is calculated using the REINFORCE algorithm to maximize the expected return. The model parameters are updated using the calculated gradients, and the learning rate α is usually set to 0.001-0.01. The AMA-UNet model is fine-tuned using the updated parameters, with special attention paid to the adaptive multi-scale feature extraction module and the dynamic attention mechanism. This process allows the system to continuously optimize the configuration strategy based on user feedback, while improving the accuracy of image processing and feature extraction by fine-tuning the AMA-UNet model.

[0139] 8. Verify the optimized configuration result. If the preset threshold is not reached, return to the step of generating AR guidance information for iterative optimization. In order to comprehensively evaluate the configuration effect, we introduce a multi-objective evaluation function (MOE):

[0140] S=w1·C+w2·F+w3·U

[0141] Among them, S is the overall score, C is the fit score, F is the functionality score, U is the user comfort score, w1, w2, and w3 are weight factors, and their sum is 1.

[0142] A typical weighting setup might be:

[0143] w1=0.4(fitting degree)

[0144] w2=0.3(functionality)

[0145] w3=0.3(user comfort)

[0146] If S is lower than a preset threshold (e.g., 0.8), the system returns to step 5, regenerates AR guidance information, and performs the next round of optimization. Through this iterative optimization process, the system can continuously improve the configuration effect until a satisfactory result is achieved or a predetermined number of iterations (e.g., a maximum of 5 iterations) is completed.

[0147] The data flow and closed-loop implementation of the present invention can be summarized as follows:

[0148] 1) Data acquisition: A high-definition camera (such as the Sony IMX766 sensor) captures 4K resolution images of the user’s ears and headphones.

[0149] 2) Image preprocessing: The image processor (such as Qualcomm Snapdragon series) executes the Multi-Scale Adaptive Contrast Enhancement (MS-ACE) algorithm.

[0150] 3) AMA-UNet processing: The preprocessed image is input into the AMA-UNet model. The model runs on a low-power AI accelerator (such as Google Edge TPU) and performs operations such as adaptive multi-scale feature extraction and dynamic attention mechanism.

[0151] 4) Feature Analysis: The feature analyzer processes the output of AMA-UNet using the improved multimodal feature fusion algorithm (IMM-FF) and is executed on the CPU.

[0152] 5) AR rendering: Based on the feature analysis results, the AR renderer (such as using ARM Mali GPU) generates a dynamic vector field and creates AR guidance information.

[0153] 6) Display and interaction: AR guidance information is presented on a high-resolution OLED display. Users interact through touch screen, voice commands or gestures, and these inputs are processed by the user interface module.

[0154] 7) Evaluation and Optimization: The evaluation module calculates the score of the multi-objective evaluation function (MOE). The reinforcement learning optimizer updates the configuration policy using the policy gradient algorithm based on this score and user feedback.

[0155] 8) Model update: The optimized parameters are used to fine-tune the AMA-UNet model. This process is performed in the background with low priority to avoid affecting real-time performance.

[0156] 9) Closed-loop iteration: If the evaluation score does not reach the threshold, the system returns to step 3 and regenerates AR guidance using the updated model. This process continues until a satisfactory result is achieved or the predetermined number of iterations is completed.

[0157] The hardware configuration of the present invention includes a main processor, Qualcomm Snapdragon 8cxGen3 (for overall control and complex calculations), an AI accelerator source Google Edge TPU (for AMA-UNet model reasoning), a GPU only ARM Mali-G78 (for AR rendering), 8GB LPDDR5 memory (to ensure smooth multi-tasking), and 128GB UFS3.1 storage (for storing models, user data and system logs).

[0158] This design forms a complete closed-loop system that continuously learns and adapts to the user's needs and preferences while maintaining efficient real-time performance. The hardware selection takes into account both performance and energy efficiency, ensuring feasibility in portable devices such as charging boxes. In this way, the present invention realizes a highly intelligent, adaptive and user-friendly headphone configuration system.

[0159] The AR-assisted earphone configuration method for the charging box screen proposed in this invention realizes high-precision and personalized earphone configuration guidance through the innovative AMA-UNet model and a series of optimization algorithms. This method not only significantly improves the user experience, but also opens up a new direction for human-computer interaction of smart wearable devices. With the further development and optimization of technology, this method has the potential to become the standard for smart device configuration in the future.

[0160] In order to verify the effectiveness of the present invention, the applicant conducted the following control experiments:

[0161] Example 1 (present invention): AR-assisted headset configuration method for charging box screen, using AMA-UNet model

[0162] Comparative Example 1: Traditional manual configuration method without AR assistance

[0163] Comparative Example 2: Basic AR assisted configuration method, using the standard U-Net model

[0164] Comparative Example 3: AR assisted configuration method, using the improved U-Net model but without adaptive multi-scale feature extraction and adversarial knowledge distillation

[0165] Test indicators and test methods:

[0166] 1. Configuration accuracy: Use 3D scanning technology to measure the position deviation of the earphone on the ear in millimeters (mm).

[0167] Detection method: Use a high-precision 3D scanner to scan the configured earphone position and compare it with the ideal position.

[0168] 2. Configuration time: The time from the start of configuration to completion, in seconds (s).

[0169] Detection method: Use a timer to record the time of the entire configuration process.

[0170] 3. User satisfaction: A 5-point Likert scale (1-5 points) was used to evaluate the user's satisfaction with the configuration process and results.

[0171] Detection method: Collect user feedback through questionnaires.

[0172] 4. Model adaptability: Tests how quickly the model adapts to new types of headphones, measured in training rounds.

[0173] Detection method: Record the number of training rounds required for the model to reach 90% configuration accuracy.

[0174] 5. Computational efficiency: The frame rate of the model running on the charging box, in frames per second (FPS).

[0175] Detection method: Use performance analysis tools to measure the running frame rate of the model on the charging box hardware.

[0176] The test results are shown in the following table:

[0177] index Example 1 Comparative Example 1 Comparative Example 2 Comparative Example 3 Configuration accuracy (mm) 0.5 3.2 1.8 1.2 Configuration time(s) 25 120 60 45 Customer satisfaction 4.8 3.2 4 4.3 Model adaptability 10 N / A 100 50 Computational efficiency (FPS) 30 N / A 15 22

[0178] The analysis and explanation of the above test results are as follows:

[0179] 1. Configuration accuracy: The configuration accuracy of Example 1 is significantly better than other methods, achieving a high accuracy of 0.5mm. This is due to the adaptive multi-scale feature extraction capability of the AMA-UNet model, which can more accurately capture the subtle features of the ear and earphones. The manual configuration accuracy of Comparative Example 1 is the lowest, and although Comparative Examples 2 and 3 use AR technology, their accuracy is still not as good as that of Example 1 due to limited model capabilities.

[0180] 2. Configuration time: Example 1 has the shortest configuration time, which is only 25 seconds. This shows that the AR-assisted method of the present invention combined with the advanced AMA-UNet model can quickly provide accurate configuration guidance. The manual configuration of Comparative Example 1 takes the longest time, and although Comparative Examples 2 and 3 use AR technology, they may require multiple adjustments due to the low model accuracy, so they take a long time.

[0181] 3. User satisfaction: Example 1 received the highest user satisfaction score (4.8 points). This shows that the present invention not only performs well in technical indicators, but also achieves significant results in user experience. The positive feedback from users on AR assistance and fast and accurate configuration reflects the practical value of the present invention.

[0182] 4. Model adaptability: Example 1 shows excellent model adaptability, and can adapt to new models of headphones in just 10 rounds of training. This is due to the adversarial knowledge distillation technology used in the present invention, which greatly improves the generalization ability and rapid learning ability of the model. Although comparative examples 2 and 3 also have certain adaptability, they require more training rounds.

[0183] 5. Computational efficiency: Example 1 can still achieve a high frame rate of 30FPS on resource-constrained charging box hardware. This is thanks to the lightweight depth-separable convolution design in the AMA-UNet model. Comparative Examples 2 and 3 are difficult to run smoothly on the charging box due to their heavy model structures.

[0184] Example 1 (the present invention) showed significant advantages in all test indicators. In particular, the performance of the present invention far exceeded that of other methods in terms of configuration accuracy, configuration time, and model adaptability. These results fully demonstrate the effectiveness of the core innovation of the present invention, the AMA-UNet model:

[0185] 1. The adaptive multi-scale feature extraction module significantly improves the configuration accuracy.

[0186] 2. The dynamic attention mechanism and lightweight design enable the model to run efficiently on the charging box, improving configuration speed and user experience.

[0187] 3. Adversarial knowledge distillation technology greatly improves the adaptability of the model, enabling it to quickly adapt to new models of headphones.

[0188] These innovations not only demonstrate advantages in technical indicators, but are also recognized by users in practical applications, as shown by user satisfaction scores. This invention successfully combines advanced deep learning technology with practical application requirements, providing an efficient, accurate and user-friendly solution to the configuration problem of smart wearable devices.

[0189] Based on the above test results, Example 1 is determined to be the best example. It performs well in all key indicators, especially in terms of configuration accuracy, speed, and user experience. This example fully reflects the core innovation of the present invention and sets a new standard for configuration technology of future smart wearable devices.

[0190] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for configuring an AR-assisted headset on a charging box screen, characterized in that: The following steps are involved: Obtain images of ears and headphones; Preprocessing the image; Processing the preprocessed image using an adaptive multi-scale attention U-Net model; Perform feature analysis and matching based on the output of the AMA-UNet model; Generate AR guidance information according to the result of the feature analysis and matching; Displaying the AR guidance information on the charging box screen and receiving user feedback on the adjustment of the earphone position; Based on the user feedback, optimizing the configuration strategy and updating the AMA-UNet model using a reinforcement learning algorithm; The optimized configuration result is verified. If the preset threshold is not reached, the step of generating AR guidance information is returned to perform iterative optimization until the preset threshold is reached or the predetermined number of iterations is completed.

2. The method according to claim 1, characterized in that The step of obtaining the image of the ear and the earphone comprises: Use image sensors to capture images at 4K resolution; The exposure parameters are adjusted using an adaptive exposure control algorithm, wherein the adaptive exposure control algorithm is defined as: E=f(L,S,N) f(L,S,N)=α·log(L)+β·S-γ·N 2 Among them, E is the exposure value, L is the ambient light intensity, S is the sensor sensitivity, N is the noise level, and α, β, and γ are adaptive parameters.

3. The method according to claim 1, characterized in that The step of preprocessing the image adopts a multi-scale adaptive contrast enhancement algorithm, which is defined as: I out =G(I in ,s,r) G(I,s,r)=I+r·(IF(I,s)) Among them, I in and I out are the input and output images respectively, s is the scale factor, r is the enhancement strength, and F(I,s) is the Gaussian blur function.

4. The method according to claim 1, characterized in that: The AMA-UNet model includes: Adaptive multi-scale feature extraction module; dynamic attention mechanism; lightweight depth-separable convolution and adaptive knowledge distillation module.

5. The method according to claim 4, characterized in that The adaptive multi-scale feature extraction module adopts a dynamic convolution kernel generation network, which is defined as: K i =DKGN(X,i) F ms =∑(w i ·Conv(X,K i )),i∈{1,3,5,7} Among them, DKGN(X,i) dynamically generates the convolution kernel according to the input X and the target kernel size i, w i is the weight.

6. The method according to claim 4, characterized in that The dynamic attention mechanism introduces the spatiotemporal attention mechanism, which is defined as: A=σ(f c (X)·f s (X)·f t (X)) F att =A·X Among them, f c ,f s ,f t are channel, spatial and temporal attention functions respectively, and σ is the sigmoid function.

7. The method according to claim 4, characterized in that The lightweight depth-wise separable convolution introduces an adaptive activation function, which is defined as: F dw =AAF(DWConv(X)) F pw =AAF(PWConv(F dw )) AAF(x)=α·x·sigmoid(β·x)+γ·x Among them, DWConv and PWConv are depth convolution and point-by-point convolution respectively, and α, β, and γ are learnable parameters.

8. The method according to claim 4, characterized in that The adaptive knowledge distillation module adopts adversarial knowledge distillation, which is defined as: L kd =α·L ce +(1-α)·D KL (S(x),T(x))0+λ·L adv (S(x),T(x)) L adv =E[log(D(T(x)))]+E[log(1-D(S(x)))] Among them, L ce is the cross entropy loss, D KL is the KL divergence, S(x) and T(x) are the outputs of the student model and the teacher model respectively, D is the discriminator, and λ is the weight factor.

9. The method according to claim 1, characterized in that: The feature analysis and matching step adopts an improved multimodal feature fusion algorithm, which is defined as: F=Φ(F visual ,F geometric ,F semantic )Φ(V,G,S)=σ(W v ·V+W q ·G+W s ·S) Among them, F v isual,F geometric ,F semantic are visual, geometric and semantic features respectively, and W v ,W g ,W s is a learnable weight matrix; the feature matching adopts an improved RANSAC algorithm and introduces an adaptive threshold: θ=μ·exp(-λ·t) Among them, θ is the matching threshold, t is the number of iterations, and μ and λ are adjustable parameters.

10. The method according to claim 1, characterized in that The step of generating AR guidance information adopts a dynamic vector field generation algorithm, which is defined as: Among them, V(x,y) is a 2D vector field, ψ(x,y) is a scalar potential field, φ(x,y) is a vector potential field, and curl represent the gradient and curl operators respectively; the optimization configuration strategy adopts the meta-reinforcement learning algorithm, where θ is the model parameter, α, β are the learning rates, L is the loss function, and f θ A multi-objective evaluation function is introduced for the step of verifying the configuration result described in the parameterized policy function, which is defined as: S=w1·C+w2·F+w3·U Among them, S is the overall score, C is the fit score, F is the functionality score, U is the user comfort score, and w1, w2, and w3 are weight factors.