Depth feature adaptive fusion method based on dynamic routing

Abstract

Aiming at the technical problems of high feature redundancy, large calculation delay, insufficient generalization performance and the like in the existing depth feature fusion technology, the invention provides a depth feature adaptive fusion method based on dynamic routing. According to the method, a multi-scale feature extraction network is constructed to obtain features, a feature routing weight calculation unit is used to realize weight matrix adaptive learning, a softmax function is used to dynamically calculate a feature routing weight, an adaptive feature fusion unit is combined to execute optimization fusion, and a multi-task joint loss function is used to guide model training. Experiments show that compared with the prior art, the scheme of the invention has the advantages that the feature extraction accuracy is improved to 99.997%, the calculation delay is reduced by 45.12%, the feature redundancy is reduced by 73.12%, the model parameter quantity is reduced by 45.09%, and the generalization performance score is improved by 6.91%. The method is suitable for depth feature fusion application in a complex scene, and has remarkable technical effects and practical values.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a deep feature adaptive fusion method based on dynamic routing. Background Technology

[0002] With the rapid development of artificial intelligence and deep learning technologies, feature extraction and fusion techniques based on deep learning have been widely applied in many fields such as computer vision, natural language processing, and speech recognition. Deep feature fusion, as a key technology in deep learning, plays a crucial role in improving model performance and optimizing computational efficiency. However, traditional feature fusion methods still face many challenges and limitations in practical applications.

[0003] In early research, researchers primarily employed simple feature concatenation or weighted summation methods for feature fusion. For example, the Multi-Scale Feature Fusion method proposed in 2018 fused features at different scales using fixed weights. While this improved model performance to some extent, the fixed weights prevented dynamic adjustment based on the characteristics of the input data, resulting in insufficient adaptability of the feature fusion. Subsequently, the Attention-based Feature Fusion technique proposed in 2019 introduced an attention mechanism to achieve adaptive feature fusion; however, this method has high computational complexity and is difficult to meet the needs of applications with high real-time requirements.

[0004] In recent years, with the continuous development of deep learning models, feature fusion technology has also been evolving towards greater intelligence and efficiency. The Dynamic Feature Aggregation method proposed in 2020 guides the feature fusion process by learning the correlations between features, but this method failed to effectively solve the feature redundancy problem, leading to a waste of computational resources. In 2021, researchers proposed a feature fusion method based on graph neural networks. Although this improved the feature representation capability, the model has a large number of parameters, and the computational overhead during training and inference remains high.

[0005] Currently, existing feature fusion techniques suffer from several problems: First, the feature extraction process suffers from high redundancy; a large number of invalid or duplicate features not only increase computational burden but may also affect model performance. Second, the computational latency in the feature fusion process is significant, making it difficult to meet the needs of real-time applications. Third, existing fusion methods lack effective adaptive mechanisms, failing to dynamically adjust the fusion strategy based on the characteristics of the input data, resulting in insufficient generalization performance of the model. Furthermore, most feature fusion methods employ fixed network structures and fusion strategies, making it difficult to adapt to the needs of different application scenarios.

[0006] To address these issues, researchers have explored various improvement schemes. For example, the AdaptiveFeature Selection method proposed in 2022 reduces feature redundancy by introducing sparsity constraints, but this method may lose important information during feature selection. The Efficient Feature Fusion framework proposed in 2023 reduces computational complexity by designing lightweight fusion modules, but the model's expressive power is somewhat limited. While these improvements have made progress in some aspects, they still fail to comprehensively solve the core problems faced by feature fusion technology.

[0007] In particular, in complex application scenarios, existing feature fusion techniques often fail to simultaneously balance computational efficiency and model performance. For example, in scenarios requiring the processing of high-dimensional features, traditional fusion methods may lead to significant computational delays; in environments with varying data distributions, fixed fusion strategies struggle to maintain stable performance. These issues severely restrict the promotion and application of deep feature fusion techniques in practical applications.

[0008] Therefore, a novel feature fusion method is urgently needed that can effectively reduce feature redundancy and computational latency while ensuring model performance, and possess good adaptability and generalization performance. This method should be able to dynamically adjust the fusion strategy according to the characteristics of the input data, maintain stable performance in different application scenarios, and have low computational complexity and resource consumption. This is of great significance for promoting the practical application and development of deep learning technology. Summary of the Invention

[0009] In view of the technical problems existing in deep feature fusion technology, such as high feature redundancy, large computational latency and insufficient generalization performance, this invention is proposed.

[0010] This invention provides a deep feature adaptive fusion method based on dynamic routing, which aims to solve the problems of efficient feature extraction, accurate calculation of dynamic routing weights, adaptive feature fusion, and model training optimization.

[0011] To address the aforementioned technical problems, this invention provides the following technical solution: a deep feature adaptive fusion method based on dynamic routing, comprising:

[0012] A multi-scale feature extraction network is constructed to extract features from the input data; a feature routing weight calculation unit is established to achieve adaptive learning of the weight matrix; an adaptive feature fusion unit is set up to perform feature fusion optimization; and a multi-task joint loss function is constructed to guide model training.

[0013] As a preferred embodiment of the deep feature adaptive fusion method based on dynamic routing described in this invention, the multi-scale feature extraction network includes a feature dimension normalization layer and a feature preprocessing unit, the mathematical expression of which is:

[0014] F(x)=σ(W x X+b x )

[0015] Where F(x) represents the output feature vector after feature preprocessing, which is essentially the result of nonlinear transformation and feature reconstruction of the original input features; the input feature X represents the feature vector in the original multidimensional feature space, which can come from the initial input of the multi-scale feature extraction network; the weight matrix W x It is a learnable parameter tensor that adaptively adjusts during model training via backpropagation; its dimension and structure determine the complexity of the feature mapping; the bias term b x The activation function σ (usually a non-linear activation function such as ReLU or Sigmoid) is used to adjust the overall offset of the feature mapping and enhance the expressive power of the model. The role of the activation function σ is to introduce non-linear transformation, enabling the network to learn more complex feature representations, thereby breaking through the limitations of linear models. This design enables the feature preprocessing unit to effectively capture the deep features of the input data and provide high-quality feature representations for subsequent dynamic routing and feature fusion stages. It is the key mathematical mechanism for achieving efficient feature extraction in the technical solution of this invention.

[0016] As a preferred embodiment of the deep feature adaptive fusion method based on dynamic routing described in this invention, the feature routing weight calculation unit calculates the feature routing weight using the following formula:

[0017]

[0018] Among them, g i (x) represents the routing weight for the i-th feature, which is achieved by weighting multiple features and then performing softmax normalization. Specifically, x represents the input data, M represents the total number of features, and W... ij It is the routing weight coefficient connecting the i-th output feature and the j-th input feature, reflecting the correlation strength and importance between different features, f j (x) represents the table of the j-th input feature, describing the feature vector in the original feature space, d kThe dimension of the feature space is represented by normalizing the feature vectors (dividing by the square root of the feature dimension), which can effectively alleviate the differences in feature scale and prevent the gradient from being too large or too small. The application of the softmax function ensures the non-negativity and probability distribution characteristics of the routing weights, so that the weights of different features can adaptively reflect their relative importance to the target task, thereby realizing a more refined and dynamic feature fusion strategy. This innovative routing weight calculation method can not only effectively capture the complex relationships between features, but also dynamically adjust the importance weights of features according to the specific features of the input data, significantly improving the expressive power and generalization performance of deep learning models in the feature extraction and fusion stages.

[0019] As a preferred embodiment of the deep feature adaptive fusion method based on dynamic routing described in this invention, the fusion process of the adaptive feature fusion unit is expressed as follows:

[0020]

[0021] Among them, F out The final fused feature vector is the combined result of multiple features weighted by dynamic routing; N represents the total number of features involved in the fusion, reflecting the dimension of multi-feature collaborative processing; g i (x) is the dynamic routing weight corresponding to the i-th feature, calculated using the softmax function. Its value dynamically reflects the importance and contribution weight of this feature under the current input x; F i (x) represents the i-th original feature vector, which carries the feature representation of the input data in a specific dimension; summation operation It embodies a weighted aggregation process for all features, enabling the final output to adaptively capture the complex feature representation of the input data. It realizes dynamic routing and adaptive weight allocation in the feature fusion process, effectively solving the key technical problems of fixed weights and insufficient adaptability in traditional feature fusion methods.

[0022] As a preferred embodiment of the deep feature adaptive fusion method based on dynamic routing described in this invention, the multi-task joint loss function is expressed as:

[0023] L = L task +α·L diversity +β·L sparse

[0024] Among them, L task L represents the main loss term for a specific task, used to measure the model's performance on the target task; diversity Represents diversity loss, aiming to enhance the richness and discriminativeness of feature representation; L sparseThe sparsity loss is used to control the sparsity of features and reduce redundant information; α and β are adjustable weighting coefficients used to balance the contributions of different loss terms. Through this multi-task joint optimization approach, this invention can effectively improve the diversity and sparsity of features while ensuring model performance, thereby achieving a more efficient and robust feature fusion process. This loss function design fully considers multiple optimization objectives under complex application scenarios, providing strong theoretical support and practical guidance for deep feature adaptive fusion technology.

[0025] The deep feature adaptive fusion method based on dynamic routing is characterized by: a feature dimensionality reduction unit for reducing feature dimensionality; a feature enhancement module for enhancing feature expressive power; a feature selection mechanism for selecting effective features; a feature importance evaluation unit for evaluating feature importance; a feature redundancy elimination mechanism for eliminating redundant features; an adaptive threshold adjustment unit for dynamically adjusting feature selection thresholds; a computation acceleration unit for improving computational efficiency; a resource scheduling optimizer for optimizing resource allocation; and a parallel processing mechanism for achieving parallel computation.

[0026] A computer device includes a memory and a processor, the memory storing a computer program, characterized in that: when the processor executes the computer program, it implements the steps of any of the following methods for the adaptive fusion of deep features based on dynamic routing.

[0027] A computer-readable storage medium having a computer program stored thereon, characterized in that: when the computer program is executed by a processor, it implements the steps of any of the deep feature adaptive fusion methods based on dynamic routing.

[0028] The beneficial effects of this invention are: feature extraction accuracy is increased to 99.997%, computation latency is reduced by 45.12%, feature redundancy is reduced by 73.12%, the number of model parameters is reduced by 45.09%, and the generalization performance score is improved by 6.91%. These improvements fully demonstrate the innovation and effectiveness of this invention in solving key technical problems in the field of deep feature fusion.

[0029] The technical solution of this invention is not only applicable to standard feature fusion scenarios, but also maintains stable performance in complex application environments. By combining a dynamic routing mechanism and an adaptive fusion strategy, this invention achieves efficient collaboration between feature extraction, route calculation, and fusion optimization, providing a new research direction and practical reference for the development of deep feature fusion technology. Attached Figure Description

[0030] This invention includes five accompanying drawings, which are used to describe the technical solution of this invention in detail.

[0031] Figure 1This is an overall architecture diagram of the deep feature adaptive fusion system based on dynamic routing of the present invention. The diagram shows the four core modules of the system: feature extraction module, dynamic routing module, feature fusion module and loss function module, as well as the data flow relationship between them.

[0032] Figure 2 This is a diagram of the multi-scale feature extraction network structure of the present invention. The diagram describes the internal structure of the feature dimension normalization layer and the feature preprocessing unit, and shows the data processing flow in the feature extraction process, including key steps such as input feature preprocessing, dimension normalization and feature enhancement.

[0033] Figure 3 This is a flowchart of the feature routing weight calculation process of the present invention. The flowchart shows in detail the calculation process of dynamic routing weight, including steps such as feature importance evaluation, routing weight coefficient calculation and softmax normalization. It also shows the adaptive learning mechanism of the weight matrix.

[0034] Figure 4 This is a schematic diagram of the adaptive feature fusion of the present invention. The diagram illustrates the working principle of the feature fusion unit, shows how to adaptively fuse different features through dynamic routing weights, and the specific implementation of the feature redundancy elimination mechanism.

[0035] Figure 5 This is a flowchart of the multi-task joint optimization process of the present invention. The flowchart illustrates the multi-task loss function optimization strategy during model training, including the calculation methods of task loss, diversity loss and sparsity loss, as well as the weight balancing mechanism among them. Detailed Implementation

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0039] Example 1 is the first embodiment of the present invention, which provides a deep feature adaptive fusion method based on dynamic routing, including:

[0040] S1: Multi-scale feature extraction

[0041] S1.1: The system first receives input data, and then... Figure 2 The multi-scale feature extraction network shown performs initial feature extraction. This network employs a feature dimension normalization layer to standardize the input data X, calculated as X_norm = (X - μ) / σ, where μ is the feature mean and σ is the standard deviation. This step ensures that the numerical distribution of features across different dimensions tends to be consistent, providing a stable foundation for subsequent processing.

[0042] S1.2: The standardized data is processed by a feature preprocessing unit, which uses the mathematical expression F(x) = σ(WxX + bx) to perform the initial feature transformation. Here, Wx is a 64×128-dimensional weight matrix, bx is a 64-dimensional bias vector, and σ uses the ReLU activation function. The preprocessing unit is constructed using a three-layer fully connected network, with output dimensions of 128, 96, and 64 for each layer, ensuring the gradual extraction and dimensionality reduction of feature information.

[0043] S1.3: To enhance the expressive power of features, the system includes a feature enhancement module. This module employs a residual connection structure to fuse the original features with the transformed features, using the fusion method F_enhanced = F(x) + X_norm. This design preserves the original feature information while introducing nonlinear transformations to improve the expressive power of the features.

[0044] S2: Feature routing weight calculation

[0045] S2.1: According to Figure 3 The feature routing weight calculation process shown begins with the system calculating the correlation matrix between features. For any two feature vectors fi and fj, their correlation coefficient is calculated using cosine similarity: sim(fi,fj)=(fi·fj) / (||fi||||fj||). This step establishes a measure of the strength of the association between features.

[0046] S2.2: Based on the correlation matrix, the system further calculates the feature routing weights. The mathematical model g_i(x) = softmax(∑(j=1 to M)w_ij·f_j(x) / √d_k) is used to dynamically allocate the weights. Here, w_ij is initialized as a random value from a standard normal distribution and is continuously updated during training. d_k is set to the feature dimension of 64, and M is the number of features, typically taking a value of 8 or 16.

[0047] S2.3: To ensure the rationality of weight allocation, the system introduces an attention mechanism. An attention score is calculated for each feature vector: attention_score = tanh(W_a·fi + b_a), where W_a is a 32×64 weight matrix and b_a is a 32-dimensional bias vector. The attention score is multiplied by the routing weights to highlight important features.

[0048] S3: Adaptive Feature Fusion

[0049] S3.1: Based on Figure 4 The diagram illustrates adaptive feature fusion, where the system constructs a feature fusion unit. This unit first aligns the dimensions of the routed features to ensure that all feature vectors have the same dimension. For features with inconsistent dimensions, adjustments are made using 1×1 convolutions, with the number of convolution kernels set to the target dimension size.

[0050] S3.2: The system uses a weighted summation method to achieve feature fusion, specifically expressed as F_out = ∑(i = 1 to N)g_i(x)·F_i(x). Here, g_i(x) is the routing weight of the i-th feature, F_i(x) is the corresponding feature vector, and N is the number of features participating in the fusion. This fusion method can adaptively adjust the contribution of features according to their importance.

[0051] S3.3: To reduce feature redundancy, the system employs a feature redundancy elimination mechanism. Principal component analysis (PCA) is performed on the fused features, retaining principal components that explain more than 95% of the variance. Simultaneously, L1 regularization is used to constrain the fusion weights, causing the weights of some less important features to approach zero.

[0052] S4: Multi-task joint optimization

[0053] S4.1: Based on Figure 5 The multi-task joint optimization process shown in the figure constructs a comprehensive loss function L = L_task + α·L_diversity + β·L_sparse. Here, L_task uses cross-entropy loss to measure the model's classification accuracy. α and β are set to 0.1 and 0.01 respectively to balance the contributions of different loss terms.

[0054] S4.2: The diversity loss L_diversity is achieved by calculating the cosine similarity matrix between feature vectors. For N feature vectors, an N×N similarity matrix S is constructed, and the diversity loss is defined as L_diversity = ||SI||_F, where I is the identity matrix and ||·||_F represents the Frobenius norm. This design encourages different features to remain independent.

[0055] S4.3: The sparsity loss L_sparse uses the L1 norm of the routing weights, expressed as L_sparse=∑|w_ij|. By controlling the sparsity of the weights, the system can automatically select the most important features to participate in the fusion, reducing computational complexity.

[0056] S5: Performance Optimization Strategies

[0057] S5.1: The system employs a computational acceleration unit to optimize processing efficiency. Batch processing technology is used in the feature extraction stage, processing 64 samples per batch. Simultaneously, the forward propagation process of the feature extraction network is computed in parallel on the GPU, significantly improving processing speed.

[0058] S5.2: In terms of resource scheduling, the system implements a dynamic load balancing mechanism. GPU memory allocation is dynamically adjusted based on the computational load of feature extraction and fusion. For large-scale feature computations, a block processing strategy is adopted, with each block size set to 1024×64 to avoid memory overflow.

[0059] S5.3: To improve training efficiency, the system implements a gradient accumulation mechanism. Model parameters are updated every four mini-batches, effectively reducing GPU memory usage. Simultaneously, a mixed-precision training strategy is employed, converting some computational processes to FP16 format, improving computation speed while maintaining accuracy.

[0060] Example 2 is the second embodiment of the present invention, which provides a deep feature adaptive fusion method based on dynamic routing, including:

[0061] 1. Experimental Environment

[0062] This experiment was conducted in a standardized high-performance computing environment, employing dual Intel Xeon 8280 CPUs and eight NVIDIA Tesla V100 GPUs. The server was equipped with 512GB of DDR4 memory and ran Ubuntu 20.04LTS. The deep learning framework used was PyTorch version 1.9.0, with CUDA version 11.3. All experiments were performed on the same hardware to ensure the comparability and reproducibility of the results. The system ambient temperature was maintained at 22±1℃, and the relative humidity was maintained within the range of 45%±5%.

[0063] Explanation of Comparison Option 1

[0064] In contrast, Scheme 1 employs a traditional feature fusion method combining CNN and attention mechanisms. This scheme uses ResNet-50 as the backbone network to extract features and weights them through an SE (Squeeze-and-Excitation) attention module. Feature fusion uses a fixed-weight strategy, with weight coefficients determined through pre-training and remaining unchanged during the inference phase. The main limitation of this scheme is that the feature fusion weights lack dynamic adjustment capabilities and cannot adaptively adjust based on the input data features.

[0065] Comparison of Scheme 2

[0066] Comparison Scheme 2 uses Graph Convolutional Networks (GCNs) for feature fusion. This scheme represents features at different levels as nodes in a graph structure and establishes relationships between features through graph convolution operations. It employs two graph convolutional layers with a hidden layer dimension of 256 and uses the ReLU activation function. While this scheme can establish relationships between features, it lacks a dynamic routing mechanism and cannot achieve adaptive feature selection and fusion.

[0067] 2. Experimental Procedure

[0068] The experiment first standardized the input data, scaling pixel values ​​to the [-1, 1] range, and performed data augmentation such as random horizontal flipping and cropping. Then, a multi-scale feature extraction network was used to obtain feature representations at different levels, with feature dimensions of 64, 128, 256, and 512. The feature preprocessing unit normalized the dimensions of the extracted features to ensure scale consistency across different levels. Next, the dynamic routing module calculated feature routing weights, using the Adam optimizer with a learning rate of 0.001 and a batch size of 32 for weight updates. In the feature fusion stage, adaptive feature fusion was performed based on the calculated routing weights, employing a feature redundancy elimination mechanism with a threshold of 0.3. Finally, a multi-task joint loss function guided model training, with a task loss weight of 1.0, a diversity loss weight of 0.5, and a sparsity loss weight of 0.3.

[0069] 3. Test methods and standards

[0070] Feature extraction accuracy was evaluated using a standard test set containing 50,000 test images covering 100 categories. Accuracy was measured by the consistency between the model's predictions and the ground truth labels. Computational latency was measured using the NVIDIA profiler tool, calculating the average processing time during the model's inference phase. Feature redundancy was evaluated by calculating the cosine similarity matrix between feature vectors and counting the number of feature pairs with a similarity exceeding a threshold of 0.8. The number of model parameters was determined using PyTorch's built-in tools, counting the total number of trainable parameters in the network. Generalization performance was evaluated under five different scenarios, including datasets with varying lighting conditions, shooting angles, resolutions, noise levels, and occlusion degrees.

[0071] 4. Experimental Results Table

[0072]

[0073] 5. Interpretation of Experimental Results

[0074] Through detailed experimental comparison and analysis, the proposed solution achieves a feature extraction accuracy of 99.997%, a 1.152 percentage point improvement compared to the 98.845% of the comparative solution 2. This is attributed to the optimized design of the multi-scale feature extraction network and the role of the feature dimension normalization layer. Regarding computational latency, the proposed solution's processing time is only 4.5ms, a 45.12% reduction compared to the 8.2ms of the comparative solution 2. This is mainly due to the efficient implementation of the dynamic routing mechanism and the optimization of the computation acceleration unit. In terms of feature redundancy, the proposed solution reduces it to 6.8%, a 73.12% reduction compared to the 25.3% of the comparative solution 2, fully validating the effectiveness of the feature redundancy elimination mechanism. Regarding the number of model parameters, the proposed solution requires only 35.8M parameters, a 45.09% reduction compared to the 65.2M of the comparative solution 2. This is attributed to the design of the feature dimensionality reduction unit and the routing optimizer. GPU memory usage is reduced from 12.3GB in the comparative solution 2 to 8.2GB, a 33.33% reduction, demonstrating the effectiveness of the resource scheduling optimizer. In terms of generalization performance, the score of the present invention reached 0.975, which is 6.91% higher than the score of 0.912 of the comparative scheme 2, proving that the present invention has stronger adaptability in different application scenarios.

[0075] Example 3 is the third embodiment of the present invention, which provides a deep feature adaptive fusion method based on dynamic routing, including:

[0076] This embodiment details the extended application scheme of the present invention in complex and ever-changing scenarios. Based on the architecture of Embodiments 1 and 2, this embodiment further enhances the system's adaptability and performance by introducing new optimization modules and adaptation mechanisms.

[0077] First, in the feature extraction module, this embodiment employs an improved multi-scale feature pyramid network structure. This network adds an adaptive feature enhancement unit to the original feature dimension normalization layer. This unit achieves multi-scale feature extraction and enhancement through depthwise separable convolution and channel attention mechanisms. Specifically, for the input feature X, it first passes through three parallel depthwise separable convolution branches with kernel sizes of 3×3, 5×5, and 7×7, respectively. Then, the channel attention module adaptively weights the features at different scales. This design can better capture feature information at different scales, improving the robustness of feature extraction.

[0078] Regarding the dynamic routing mechanism, this embodiment introduces a routing weight calculation method based on graph attention networks. This method models the relationships between features as a fully connected graph structure and calculates the association strength between feature nodes through a multi-head attention mechanism. Specifically, for each feature node, its attention score with all other nodes is first calculated, and then normalized using a softmax function to obtain the final routing weight. This graph-based dynamic routing mechanism can better capture the complex dependencies between features.

[0079] The feature fusion module employs an improved adaptive gating mechanism. This mechanism uses a two-stream network structure to learn importance weights and complementarity weights for features separately. The main stream network learns the importance score for each feature, while the auxiliary stream network focuses on the complementarity relationships between features. The outputs of the two branches are fused through an adaptive gating unit that dynamically adjusts the contribution ratio of the two branches according to the needs of the current task. This design not only improves the accuracy of feature fusion but also effectively reduces feature redundancy.

[0080] To better handle computational demands in complex scenarios, this embodiment designs a distributed computing framework. This framework employs a parameter server architecture, distributing tasks such as feature extraction, routing computation, and feature fusion across multiple computing nodes for parallel execution. The resource scheduling optimizer within the framework dynamically adjusts the allocation strategy of computing resources based on task priority and computational load. Furthermore, by introducing an asynchronous parameter update mechanism, the communication overhead between nodes is significantly reduced.

[0081] This embodiment also optimizes the model training process. First, an improved multi-task learning strategy is adopted, adding an adversarial loss term to the original task loss, diversity loss, and sparsity loss. This loss term improves the model's robustness by generating adversarial examples. Second, a dynamic weight balancing mechanism is introduced, which adaptively adjusts the weight coefficients of each loss term based on their changing trends during training.

[0082] In terms of performance monitoring, this embodiment implements a comprehensive real-time monitoring system. This system includes a performance evaluation unit, a resource monitoring unit, and a fault detection unit. The performance evaluation unit assesses system performance in real time using multiple key indicators, the resource monitoring unit tracks the usage of computing resources, and the fault detection unit promptly detects and handles anomalies in system operation. This monitoring data is used to optimize system operating parameters and resource allocation strategies.

[0083] This embodiment also pays special attention to the model's generalization ability. By introducing a domain adaptation layer and a feature alignment module, knowledge transfer between different scenarios is achieved. The domain adaptation layer reduces the difference between the source and target domains by adjusting the feature distribution. The feature alignment module aligns the feature space by minimizing the maximum average difference, thereby improving the model's adaptability in new scenarios.

Claims

1. A deep feature adaptive fusion method based on dynamic routing, characterized in that, include: Construct a multi-scale feature extraction network to extract features from the input data; Establish a feature routing weight calculation unit to achieve adaptive learning of the weight matrix; Configure an adaptive feature fusion unit and perform feature fusion optimization; Construct a multi-task joint loss function to guide model training.

2. The deep feature adaptive fusion method based on dynamic routing as described in claim 1, characterized in that: The multi-scale feature extraction network includes a feature dimension normalization layer and a feature preprocessing unit. The mathematical expression for the feature preprocessing unit is as follows: F(x)=σ(W x X+b x ) Where F(x) represents the output feature vector after feature preprocessing; X represents the feature vector in the original multidimensional feature space; W x It is a learnable parameter tensor; b x Used to adjust the overall offset of the feature map; the activation function σ (usually a non-linear activation function such as ReLU or Sigmoid is chosen) introduces a non-linear transformation, enabling the network to learn more complex feature representations.

3. The deep feature adaptive fusion method based on dynamic routing as described in claim 2, characterized in that: The feature routing weight calculation unit calculates the feature routing weight using the following formula: Among them, g i (x) represents the routing weight for the i-th feature, which is achieved by weighting multiple features and then performing softmax normalization. x represents the input data, M represents the total number of features, and W ij f is the routing weight coefficient connecting the i-th output feature and the j-th input feature. j (x) represents the table of the j-th input feature, d k The softmax function, representing the dimension of the feature space, ensures the non-negativity and probability distribution characteristics of the routing weights.

4. The deep feature adaptive fusion method based on dynamic routing as described in claim 3, characterized in that: The fusion process of the adaptive feature fusion unit is represented as follows: Among them, F out The final fused output feature vector is represented by N; N represents the total number of features involved in the fusion; g i (x) is the dynamic routing weight corresponding to the i-th feature; F i (x) represents the i-th original feature vector; summation operation This reflects a weighted aggregation process for all features.

5. The deep feature adaptive fusion method based on dynamic routing as described in claim 4, characterized in that: The multi-task joint loss function is expressed as follows: L=L task +α·L diversity +β·L sparse Among them, L task Indicates the main loss item for a specific task; L diversity Represents a loss of diversity; L sparse The loss is sparsity loss; α and β are adjustable weighting coefficients.

6. The deep feature adaptive fusion method based on dynamic routing as described in claim 5, characterized in that: Feature dimensionality reduction unit, used to reduce feature dimensionality; The feature enhancement module is used to enhance feature representation capabilities; Feature filtering mechanism, used to filter effective features.

7. The deep feature adaptive fusion method based on dynamic routing as described in claim 6, characterized in that: Feature importance assessment unit, used to assess the degree of feature importance; Feature redundancy elimination mechanism, used to eliminate redundant features; An adaptive threshold adjustment unit is used to dynamically adjust the feature selection threshold.

8. A deep feature adaptive fusion method based on dynamic routing, employing the method described in any one of claims 1-7, characterized in that: Computational acceleration unit, used to improve computational efficiency; A resource scheduling optimizer is used to optimize resource allocation. Parallel processing mechanisms are used to achieve parallel computing.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep feature adaptive fusion method based on dynamic routing as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep feature adaptive fusion method based on dynamic routing as described in any one of claims 1 to 7.

Images