Image big data classification and identification method and system based on deep learning

Abstract

The invention relates to the field of computer vision and deep learning, and discloses an image big data classification and recognition method and system based on deep learning, and the method comprises the steps: generating a gating matrix through the extraction of an image frequency domain energy coefficient, compressing a convolution kernel through the combination of asymmetric tensor decomposition, and carrying out the self-adaptive training through cross-modal semantic alignment and a meta-learning task. Efficient classification reasoning of dynamic path selection is realized, and the precision and the calculation efficiency are improved; the system comprises a frequency domain analysis module, a dynamic sparse gating module, an asymmetric tensor decomposition module, a meta-learning task generation module, a cross-modal alignment module and a dynamic inference engine module. According to the method, through cross-modal semantic alignment and meta-learning task optimization, in combination with lightweight parameter storage and edge calculation path selection, fine-grained classification precision improvement, model compression and high-efficiency reasoning are realized, and the calculation efficiency and generalization ability in a complex scene are remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and deep learning, and specifically to a method and system for classifying and recognizing image big data based on deep learning. Background Art

[0002] With the explosive growth of image data, deep learning-based image classification technologies face multiple challenges in terms of computational efficiency, model generalization, and hardware compatibility. Traditional deep neural networks typically use fixed network structures and static computational paths when processing large-scale, high-resolution, multimodal image data. This makes it difficult to dynamically allocate computing resources based on image content characteristics, resulting in a large amount of redundant computation and energy consumption. In edge computing scenarios, in particular, the number of parameters and computational complexity of existing models often exceed the device's capacity, making it difficult to meet real-time requirements.

[0003] While existing model compression techniques can reduce the number of parameters, static compression strategies cannot adapt to input images with varying frequency domain characteristics. Images rich in high-frequency detail suffer from overcompression, leading to reduced classification accuracy, while images dominated by low-frequency features suffer from parameter redundancy. Furthermore, traditional methods, when handling fine-grained classification tasks, rely on large amounts of annotated data and manually designed features, making it difficult to effectively capture the discriminative characteristics of local object components. Furthermore, they underutilize cross-modal semantic information, limiting the model's robustness in complex scenarios.

[0004] In dynamic network design, existing solutions often achieve sparse computation through gating mechanisms. However, these gating strategies are typically based on heuristic rules or single-modal features and lack quantitative analysis of image frequency domain characteristics, resulting in a weak correlation between computational path selection and image content. Furthermore, during dynamic network training, frequent changes in sparse structures can easily lead to unstable parameter optimization and increase the risk of model overfitting. When deployed on edge devices, the computational path switching and parameter storage mechanisms of dynamic networks have yet to effectively coordinate, making it difficult to balance real-time inference efficiency with resource constraints.

[0005] Therefore, the present invention proposes a method and system for classifying and recognizing image big data based on deep learning to address the shortcomings of the existing technology. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a method and system for image big data classification and recognition based on deep learning, which solves the problems of static computational redundancy, insufficient cross-modal semantic utilization, weak fine-grained feature discrimination ability and low edge deployment efficiency when processing large-scale image data.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for classifying and recognizing image big data based on deep learning, the method comprising the following steps:

[0008] S1. Perform frequency domain transformation on the input image, extract the energy distribution of high-frequency components and low-frequency components, and generate frequency domain energy weight coefficients;

[0009] S2. For the convolution kernel weights in the deep learning network, a sparse gating matrix is ​​dynamically generated based on the frequency domain energy weight coefficients, and a gating strategy is formed by screening activated neuronal connections through learnable channel importance coefficients and adaptive thresholds;

[0010] S3. Performing asymmetric tensor decomposition on the convolution kernel weights according to the activation state of the sparse gating matrix to obtain tensor decomposition parameters, and synchronously updating the tensor decomposition parameters and the gating strategy by jointly optimizing the objective function to construct a dynamic network.

[0011] S4. Based on the input image and pre-extracted semantic segmentation information, construct a fine-grained meta-learning task set, and perform task adaptive training on the dynamic network using a second-order gradient update algorithm;

[0012] S5. Semantically aligning the feature representation of the dynamic network with the pre-generated cross-modal text description embedding vector, and constructing a distillation loss function to constrain and optimize the dynamic network;

[0013] S6. Dynamically select a calculation path to complete classification reasoning based on the frequency domain energy weight coefficient, sparse gating matrix and tensor decomposition parameters.

[0014] Preferably, in step S1, the step of generating the frequency domain energy weight coefficient includes:

[0015] For the input image Perform fast Fourier transform to obtain frequency domain representation Where H, W, and C are the height, width, and number of channels of the image, respectively. The high-frequency components are divided according to the preset frequency range. and low-frequency components The frequency domain energy weight coefficient α is generated by calculating the Frobenius norm ratio of the high-frequency component to the low-frequency component. The expression of the frequency domain energy weight coefficient α is:

[0016]

[0017] Where α is the frequency domain energy weight coefficient; ||·|| F represents the Frobenius norm of the matrix; F high is the high frequency component; F low is the low-frequency component; ∈ is a smoothing constant to prevent zero division.

[0018] Preferably, in step S2, the step of dynamically generating a sparse gating matrix includes:

[0019] Based on the frequency domain energy weight coefficient α and the learnable channel importance coefficient Generate a sparse gating matrix G∈{0,1} C×H×W , each element G (c,h,w) The activation state of is determined according to the sparse gated element rule, the formula of which is:

[0020]

[0021] Where G (c,h,w) Represents the element value at the position with channel index c, height index h, and width index w in the sparse gating matrix G; α is the frequency domain energy weight coefficient; β (c) represents the learnable importance parameter of the cth channel; τ is the dynamically adjusted threshold;

[0022] The sparse gated element rule combines the frequency domain energy weight coefficient α with the learnable channel importance coefficient Combined, it is determined whether to activate the entire feature channel c by judging whether the product of the two is greater than the dynamically adjusted activation threshold τ;

[0023] The sparse gated element rule makes the α value larger for images with rich high-frequency details, and the network tends to activate more feature channels for refined analysis; while for images with smooth content, the α value is smaller, and only the most critical channels are retained, thereby suppressing redundant calculations.

[0024] Preferably, in step S3, the steps of performing asymmetric tensor decomposition on the convolution kernel weights to obtain tensor decomposition parameters, and synchronously updating the tensor decomposition parameters and the gating strategy by jointly optimizing the objective function include:

[0025] The convolution kernel weight Decomposed into Tucker form:

[0026]

[0027] Where W is the convolution kernel weight; the core tensor × n Represents the n-module product of a tensor and a matrix; K represents the spatial dimension of the convolution kernel; K×K means that the height and width of the convolution kernel are both K; the factor matrix C out C represents the number of output channels of the convolutional layer. ln represents the number of input channels of the convolutional layer, and the rank values ​​R1, R2, and R3 are allocated and determined according to the frequency domain energy weight coefficient α according to the rank value allocation rule; wherein, the rank value allocation rule is:

[0028]

[0029] Where R max is the preset maximum rank value;

[0030] Based on the convolution kernel weight W and the objective function, joint optimization is achieved by minimizing the weight reconstruction error with sparsity constraints; wherein the objective function is:

[0031]

[0032] Where min W,G Indicates that the objective function is to minimize its value by adjusting W and G, where W is the convolution kernel weight and G represents the sparse gating matrix; ⊙ represents element-by-element multiplication; ||·|| F is the Frobenius norm of the matrix; ||G||0 represents the number of non-zero gates; λ is the sparsity control coefficient.

[0033] Preferably, in step S4, the step of constructing a fine-grained meta-learning task set includes:

[0034] Generate a semantic map M∈{0,1} for the input image using pre-trained semantic segmentation information H×W×L , where L represents the number of categories of object parts in the image, h is the height index, and W is the convolution kernel weight;

[0035] Based on the semantic graph, different parts of the same object are defined as independent categories, and a fine-grained meta-learning task set is constructed. The fine-grained meta-learning task set Include support set and queryset Meta-task set Where S k Represents the kth task T in meta-learning k The support set of (x j ,y j ) represents a sample pair in the support set, x j represents the i-th input sample, y j Represents the input sample x j Corresponding label; y i ,y j ∈{1,2,...,L} is the component category label; N represents the support set S k The number of samples in Q k Represents the kth task T in meta-learning k query set; j is the index of the sample; M represents the query set Q k The number of samples in; T represents the entire meta-learning task set; T k represents the kth meta-task; k is the index of the meta-task; K represents the meta-task.

[0036] Preferably, in step S4, the step of performing task adaptive training on the dynamic network using a second-order gradient update algorithm includes:

[0037] Based on each meta-task T k , and a second-order gradient update algorithm to update the dynamic network parameters through the inner loop learning rate η, wherein the formula of the second-order gradient update algorithm is:

[0038]

[0039] Where θ k Represents the kth meta-task T k The adjusted task-specific model parameters; θ represents the global shared parameters of the model; η is the inner loop learning rate; represents the support set loss function The gradient with respect to the global parameter θ is is the support set loss function, G k is the sparse gating matrix corresponding to the current task;

[0040] Based on the updated dynamic network parameters and the global parameter update model, the global parameters are updated by the outer loop learning rate γ, wherein the global parameter update model is:

[0041]

[0042] Where ← represents the assignment operation; γ is the outer loop learning rate; represents the gradient of the sum of all meta-task query set losses with respect to the global parameter θ, represents the sum of all meta-tasks from the 1st to the Kth, where K represents the total number of meta-tasks. is the loss function of the query set; the gradient calculation includes the second-order derivative term of the sparse gating matrix

[0043] Preferably, in step S5, the step of semantically aligning the feature representation of the dynamic network with the pre-generated cross-modal text description embedding vector includes:

[0044] The text description corresponding to the image is converted into a cross-modal text description embedding vector through a pre-trained cross-modal text encoder Where d is the embedding dimension;

[0045] Embedding vector based on cross-modal text description Mapping the output features of the dynamic sparse network into image embedding vectors Where G i ⊙W(x i ) represents the sparse gating matrix G i ∈{0,1} C×H×WThe filtered convolution output; the distillation loss function is constructed by minimizing the mean squared error between the image and text embeddings:

[0046]

[0047] Where, represents the distillation loss function; Indicates the sum of N samples in a batch or set, where N is the number of training samples and i is the index of the sample; E image (·) represents an image embedding function or network module; G i is the value corresponding to the i-th input sample x i The sparse gating matrix W(x i ) represents the input image x i The feature map or output tensor produced after the dynamic network; E text (·) represents a text embedding function or a pre-trained text encoder; T i Represents the i-th input image x i Corresponding text description; ||·||2 represents the Euclidean norm; T i is the text description of the i-th sample.

[0048] Preferably, in step S6, the step of dynamically selecting a calculation path according to the frequency domain energy weight coefficient, the sparse gating matrix and the tensor decomposition parameter includes:

[0049] Calculate the new frequency domain energy weight coefficient α based on the input image new , based on the new frequency domain energy weight coefficient α new and learnable channel importance coefficients Generate real-time sparse gating matrix G new ∈{0,1} C×H×W , wherein the real-time sparse gating matrix G new ∈{0,1} C×H×W The element activation rule in is: when α new β (c) When >τ, G new =1, where τ is the dynamically adjusted threshold;

[0050] Combined with the core tensor after asymmetric tensor decomposition and the factor matrix U (1) ,U (2) ,U (3) ,pass Reconstruct the convolution kernel weights and perform G new ⊙W decomposed The classification probability distribution is obtained by sparse calculation, where the rank values ​​R1, R2, and R3 are calculated according to Dynamic allocation, R max is the preset maximum rank value, × n The present invention also provides an image big data classification and recognition system based on deep learning, the system comprising the following modules:

[0051] A frequency domain analysis module, used for performing frequency domain transformation on the input image to generate a frequency domain energy weight coefficient;

[0052] A dynamic sparse gating module, configured to generate a sparse gating matrix based on the frequency domain energy weight coefficient and the learnable channel importance coefficient to form a gating strategy;

[0053] An asymmetric tensor decomposition module, for performing asymmetric tensor decomposition on convolution kernel weights to generate tensor decomposition parameters, and jointly optimizing them with the gating strategy to construct a dynamic network;

[0054] a meta-learning task generation module, configured to construct fine-grained meta-learning tasks based on pre-extracted semantic segmentation information and perform task-adaptive training on the dynamic network;

[0055] A cross-modal alignment module, configured to semantically align the feature representation of the dynamic network with the cross-modal text description embedding vector to constrain network optimization;

[0056] The dynamic inference engine module is used to dynamically select the calculation path to complete the classification task according to the frequency domain energy weight coefficient of the input image and the trained gating strategy and tensor decomposition parameters during inference.

[0057] Preferably, the dynamic reasoning engine module is deployed on an edge computing device and includes:

[0058] Lightweight tensor decomposition storage unit, used to store dynamically adjusted core tensors and factor matrices;

[0059] A real-time gating controller is used to execute the trained gating strategy based on the frequency-domain energy weight coefficients of the real-time input image to generate an immediate sparse gating matrix and screen the activated computational paths.

[0060] The present invention provides a method and system for classifying and recognizing image big data based on deep learning. It has the following beneficial effects:

[0061] 1. This invention uses frequency-domain energy weighting coefficients to analyze the distribution of high- and low-frequency image features in real time. Combined with a dynamic sparse gating matrix, it adaptively activates key neuronal connections and blocks redundant computational paths. This mechanism dynamically adjusts the network structure based on image content characteristics, significantly reducing computational complexity while maintaining classification accuracy. It is particularly suitable for real-time processing of massive image data.

[0062] 2. This paper introduces asymmetric tensor decomposition technology, dynamically assigning decomposition ranks based on frequency domain features and performing differentiated low-rank compression on convolution kernel weights. Combined with lightweight parameter storage and real-time reconstruction mechanisms, this effectively reduces model parameter count and memory usage, providing a hardware-friendly lightweight model architecture for edge device deployment.

[0063] 3. This paper constructs fine-grained meta-learning tasks using semantic segmentation information and combines it with a second-order gradient optimization strategy to force the network to rapidly adapt to dynamic sparse structures with a small number of samples. This design strengthens the model's ability to discriminate local features of objects (such as parts and textures), significantly improving the robustness of fine-grained classification in complex scenarios.

[0064] 4. This invention introduces global semantic supervision across language modalities by aligning the semantic distillation of text description embeddings with image features. This mechanism overcomes the limitations of a single visual modality and enhances the network's understanding of high-level image semantics through cross-modal knowledge transfer, effectively alleviating the problem of key feature loss caused by dynamic sparse computation.

[0065] 5. This invention uses a dynamic inference engine to drive computational path selection using frequency-domain features, combined with lightweight tensor decomposition storage and hardware-level sparse computing optimization to achieve low-latency, energy-efficient classification inference. This solution balances computational efficiency and accuracy, can be flexibly deployed on resource-constrained mobile terminals and edge devices, and meets the real-time processing requirements of scenarios such as the Internet of Things. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the method of the present invention;

[0067] Figure 2 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0068] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0069] See also Figure 1 , an embodiment of the present invention provides a method for classifying and recognizing image big data based on deep learning, the method comprising the following steps:

[0070] S1. Perform frequency domain transformation on the input image, extract the energy distribution of high-frequency components and low-frequency components, and generate frequency domain energy weight coefficients;

[0071] In this embodiment, the generation process of the frequency domain energy weight coefficient is achieved through the following technical solutions:

[0072] First, the input image Perform fast Fourier transform (FFT) to convert it from the spatial domain to the frequency domain to obtain a frequency domain representation in complex form Where H and W are the height and width of the input image, respectively, and C is the number of channels. The frequency domain transform uses a standard two-dimensional FFT algorithm to process each channel to separate the spectral components of different frequencies in the image.

[0073] Subsequently, the frequency domain representation F is divided into components based on a preset frequency range. Specifically, the center area of ​​the frequency domain coordinate system (for example, the low-frequency area within the radius r) is defined as the low-frequency component The edge area (high-frequency area outside the radius r) is defined as the high-frequency component The preset frequency range can be dynamically adjusted according to the image resolution. For example, the radius r of the low-frequency region can be set to 1 / 4 to 1 / 8 of the short side size of the image to adapt to image features of different scales.

[0074] After the component division is completed, the frequency domain energy weight coefficient α is generated by calculating the energy distribution ratio of the high-frequency component and the low-frequency component. Specifically, the energy of the high-frequency component is represented by F high Frobenius norm of ||F high || F Characterized by the low-frequency component energy ||F low || F To eliminate the risk of the denominator being zero, a smoothing constant ∈ (preferably 10 -6 magnitude), the final frequency domain energy weight coefficient is calculated as:

[0075]

[0076] Where α is the frequency domain energy weight coefficient; ||·|| F Represents the Frobenius norm of the matrix. The Frobenius norm is calculated as the square root of the sum of the absolute values ​​of the matrix elements, that is, for any complex matrix Its norm ∈ is a smoothing constant to prevent zero; F high is the high frequency component; F low is the low-frequency component;.

[0077] The frequency-domain energy weight coefficient α quantifies the energy difference between high-frequency details (such as edges and textures) and low-frequency background (such as smooth areas) in an image. When the image contains abundant high-frequency information, the α value increases, and vice versa. This coefficient serves as the core basis for subsequent dynamic network structure adjustments, such as controlling the sparse activation of convolution kernels and the allocation of asymmetric decomposition ranks.

[0078] Preferably, the division rule between high-frequency and low-frequency components can be dynamically optimized using learnable parameters. For example, the frequency boundary radius r is set as a differentiable function related to the image content and adjusted via gradient descent. Furthermore, the introduction of the smoothing constant ∈ ensures numerical stability, avoiding abnormal fluctuations in the weight coefficient caused by excessively low low-frequency energy.

[0079] S2. Dynamically generate a sparse gating matrix based on the frequency domain energy weight coefficient, and screen activated neuronal connections through learnable channel importance coefficients and adaptive thresholds;

[0080] In this embodiment, the generation process of the dynamic sparse gating matrix is ​​achieved through the following technical solutions:

[0081] First, based on the frequency domain energy weight coefficient α calculated in step S1, combined with the learnable channel importance parameter Construct a sparse gating matrix G∈{0,1} C×H×W Where C is the number of channels of the input image, H and W are the height and width of the feature map, respectively. The channel importance parameter β is automatically optimized through back propagation to quantify the difference in contribution of different channels to the classification task.

[0082] Specifically, each element in the sparse gating matrix G (c,h,w) The activation state is determined by the following rules:

[0083]

[0084] Where G (c,h,w) Represents the element value at the position with channel index c, height index h, and width index w in the sparse gating matrix G; α is the frequency domain energy weight coefficient; β (c) represents the learnable importance parameter of the cth channel. Its initial value can be randomly initialized using a normal distribution and updated via gradient descent during training. τ is a dynamically adjusted threshold used to control the overall sparsity rate. The adjustment strategy for the threshold τ is preferably based on a feedback mechanism based on the network activation rate. For example, when the overall activation rate of the sparse gate is lower than the preset target, τ is gradually reduced to allow more connections to be activated. Otherwise, τ is increased to enhance sparsity.

[0085] This rule combines the frequency domain energy weight coefficient α with the learnable channel importance coefficient Combined, it is determined whether to activate the entire feature channel c by judging whether the product of the two is greater than the dynamically adjusted activation threshold τ;

[0086] This rule makes the α value larger for images with rich high-frequency details, and the network tends to activate more feature channels for refined analysis; while the α value is smaller for images with smooth content, and only the most critical channels are retained, thereby suppressing redundant calculations;

[0087] The frequency-domain energy weight coefficient α acts as a global modulator in this process. For images dominated by high-frequency components (larger α), the model tends to enhance the activation of high-frequency-sensitive channels, such as those related to edge detection. Conversely, for images dominated by low-frequency components, the weight of background feature extraction channels is strengthened. Through the product of α and β, this scheme achieves dynamic recalibration of channel importance guided by frequency-domain saliency.

[0088] Furthermore, the generation of the sparse gating matrix G is spatially adaptive. For each spatial location (h, w) in the feature map, its activation state is independently calculated, enabling the model to dynamically block redundant computations based on the importance of local features. For example, in image background regions (dominated by low frequencies), the gating matrix can suppress channel activations associated with high-frequency details, thereby reducing the transfer of ineffective features.

[0089] Preferably, the optimization process of the learnable parameter β is performed in conjunction with the network classification loss function, and the gradient is propagated back to the channel importance parameter via the chain rule, enabling adaptive learning of the contribution of different channels under the regulation of frequency domain features. The adjustment process of the dynamic threshold τ can be designed as a statistical method based on a sliding window, for example, calculating the average activation rate of the current batch of data at regular training cycles and dynamically modifying τ based on the preset target activation rate.

[0090] S3. Perform asymmetric tensor decomposition on the convolution kernel weights according to the activation state of the sparse gating matrix, and synchronously update the decomposition parameters and the gating strategy by jointly optimizing the objective function;

[0091] In this embodiment, the asymmetric tensor decomposition and joint optimization are implemented through the following technical solutions:

[0092] First, based on the sparse gating matrix G∈{0,1} generated in step S2 C×H×W , for the original convolution kernel weight Perform dynamic asymmetric tensor decomposition, where K represents the spatial dimension of the convolution kernel; K×K means that the height and width of the convolution kernel are both K. Specifically, decompose the convolution kernel weight W into Tucker form:

[0093]

[0094] Where W is the convolution kernel weight; is the core tensor, is a factor matrix, which represents the low-rank mapping relationship between output channel, input channel and spatial dimension, C out C represents the number of output channels of the convolutional layer. ln is the number of input channels of the convolutional layer; n Represents the n-modular product of a tensor and a matrix; The decomposition rank values ​​R1, R2, and R3 are dynamically allocated according to the frequency domain energy weight coefficient α calculated in step S1. The specific rules are:

[0095]

[0096] Where R max The maximum rank value is preset and used to limit the magnitude of the decomposed parameters. Through this allocation rule, when high-frequency features of the image are significant (α is large), the output channel dimension retains a higher rank value to enhance the ability to express details, while the input channel and spatial dimensions have their ranks compressed to reduce redundant calculations. Conversely, when low-frequency features dominate, the output channel rank is compressed first.

[0097] Furthermore, the decomposition parameters (core tensors and the factor matrix U (1) ,U (2) ,U (3) ) is updated synchronously with the sparse gating matrix G through joint optimization of the objective function. The objective function is defined as:

[0098]

[0099] Where min W,G Indicates that the objective function is to minimize its value by adjusting W and G, where W represents the original, undecomposed convolution kernel weight tensor and G represents the sparse gating matrix; ⊙ represents element-by-element multiplication; ||·|| F is the Frobenius norm; the first term is the Frobenius norm of the weight reconstruction error, which measures the difference between the decomposed weight and the original weight; the second term ||G||0 represents the number of non-zero elements in the sparse gating matrix, and the sparsity control coefficient λ is used to adjust the overall sparsity of the network; λ is the sparsity control coefficient.

[0100] During the optimization process, the sparse gating matrix G and the tensor decomposition parameters are jointly updated via a gradient descent algorithm. Specifically, for the gating matrix G, its binary nature is used to implement gradient backpropagation via a straight-through estimator, prioritizing optimization of the decomposition parameters corresponding to non-zero gate positions. For the decomposition parameters, the gradient calculation involves chain derivatives of the factor matrix and the core tensor, ensuring that the decomposition structure can adapt to the sparsity pattern of the gating matrix. Preferably, the sparsity control coefficient λ can be dynamically decayed as the training progresses, initially strengthening the sparse constraints to screen key connections and then gradually relaxing them to stabilize parameter convergence.

[0101] The asymmetric tensor decomposition dynamically adjusts the rank value allocation through the frequency domain energy coefficient α, so that the network can adaptively allocate computing resources according to the image content characteristics. For example, for images with rich high-frequency details, the core tensor In the output channel dimension, the higher rank value is retained, and the factor matrix U (1) The dimension of the image is expanded to enhance the feature expression ability; while the low-frequency dominated image is compressed to reduce the computational complexity by compressing the output channel rank value.

[0102] In addition, the sparse gating matrix G plays a dual role in the joint optimization: on the one hand, it is the element-wise multiplication (G⊙W decomposed ) shields redundant weights, reducing forward computational effort. Furthermore, the ||G||0 constraint guides the decomposition parameters to converge toward a sparse subspace, preventing overfitting. This coupled optimization mechanism of the gating matrix and decomposition parameters achieves a dynamic balance between network structure and computational efficiency.

[0103] S4. Based on the input image and pre-extracted semantic segmentation information, construct a fine-grained meta-learning task set, and perform task adaptive training on the dynamic network using a second-order gradient update algorithm;

[0104] In this embodiment, the construction of the fine-grained meta-learning task set and the second-order gradient update process are implemented through the following technical solutions:

[0105] First, based on the input image And pre-extracted semantic segmentation information, construct a fine-grained meta-learning task set. Specifically, a pre-trained semantic segmentation model (such as DeepLabv3) is used to generate a semantic map M∈{0,1} for the input image. H×W×L , where L represents the number of categories of object parts in the image (such as "headlights", "doors", "tires", etc.). The semantic map is obtained by pixel-by-pixel classification, and the category label m for each position (h, w) is h,w ∈{1,2,...,L} represents the component category to which the pixel belongs.

[0106] Based on the semantic graph, different component regions of the same object are defined as independent pseudo categories to construct a meta-task set. Each meta-task T k Include support set and queryset Where S k Represents the kth task T in meta-learning k The support set of (x i ,y i ) represents a sample pair in the support set, x i represents the i-th input sample, y i Represents the input sample x i Corresponding label; y i ,y j ∈{1,2,...,L} is the component category label; N represents the support set S k The number of samples in Q k Represents the kth task T in meta-learning k query set; j is the index of the sample; M represents the query set Q k The number of samples in; T represents the entire meta-learning task set; T k Denotes the kth meta-task; k is the index of the meta-task; and K represents the total number of meta-tasks in the meta-task set. The samples in the support and query sets are generated by randomly sampling image patches within the same semantic region. This ensures that all samples in the same task come from different parts of the same object, forcing the model to learn fine-grained feature discrimination capabilities.

[0107] Furthermore, the dynamic network is trained for task adaptation using a second-order gradient update algorithm. For each meta-task T k , first based on the support set S k Update dynamic network parameters. Specifically, adjust the current task parameters θ through the inner loop learning rate η k :

[0108]

[0109] Where θ k Represents the kth meta-task T k The adjusted task-specific model parameters; θ represents the global shared parameters of the model; η is the inner loop learning rate; represents the support set loss function The gradient with respect to the global parameter θ is is the support set loss function, G k is the sparse gating matrix corresponding to the current task; is the support set loss function (such as cross entropy loss); G k ∈{0,1} C×H×W is the sparse gating matrix corresponding to the current task, and its generation rule is inherited from step S2 and is related to the input image content.k Maintain fixation within the task to ensure stability of task-specific structures.

[0110] Then, based on the query set Q k The loss function Update the global parameters θ through the outer loop learning rate γ:

[0111]

[0112] Where ← represents the assignment operation; γ is the outer loop learning rate; represents the gradient of the sum of all meta-task query set losses with respect to the global parameter θ, represents the sum of all meta-tasks from the 1st to the Kth, where K represents the total number of meta-tasks. is the loss function of the query set;

[0113] The gradient calculation needs to explicitly include the sparse gating matrix G k The second-order derivative of To model the coupling optimization relationship between the gating strategy and the network parameters. Preferably, the second-order derivative is implemented by an automatic differentiation tool (such as PyTorch's double-backward mechanism) to ensure the integrity of the gradient chain propagation.

[0114] The core goal of this meta-task construction and second-order optimization mechanism is to improve the model's generalization capabilities for dynamic sparse structures. By forcing the network to quickly adapt the sparse gating pattern on a small number of support set samples (inner loop) and verifying generalization performance on the query set (outer loop), the model can learn gating strategies and decomposition parameters that are robust to frequency domain feature variations. Furthermore, task division guided by semantic segmentation information allows the meta-task to focus on the fine-grained differences between different parts of the same object, thereby enhancing the network's ability to represent local features.

[0115] Preferably, the number of samples N and M in the support set and query set can be set to the same order (e.g., 5-20) to balance the data distribution between training and validation within the task. The ratio of the inner loop learning rate η to the outer loop learning rate γ can be determined by grid search, and usually γ < < η is satisfied to stabilize the meta-optimization process.

[0116] S5. Semantically aligning the feature representation of the dynamic network with the cross-modal text description embedding vector, and constructing a distillation loss function to constrain the network optimization process;

[0117] In this embodiment, the process of constructing the semantic alignment and distillation loss function is implemented through the following technical solutions:

[0118] First, the text description T corresponding to the image is converted through a pre-trained cross-modal text encoder (such as CLIP text encoder)i Convert to embedding vector Where d is the embedding dimension. The text description T i Contains keywords or phrases related to the semantics of the image (such as "the wheels of a red car"). The encoder extracts the global semantic information of the text through the self-attention mechanism and generates a dense vector representation that is highly relevant to the image content.

[0119] At the same time, the output feature G of the dynamic sparse network i ⊙W(x i ) is mapped to an image embedding vector Specifically, the output features of the dynamic network are globally averaged and then projected to the same dimension d as the text embedding through the fully connected layer, where G i ∈{0,1} C×H×W is the sparse gating matrix generated in step S2, W(x i ) is the convolution kernel weight after the asymmetric tensor decomposition in step S3. The gating matrix G i The feature channels and spatial positions are filtered to ensure that only key features related to the high-frequency or low-frequency characteristics of the current image are retained to participate in the embedding mapping.

[0120] Furthermore, the distillation loss function is constructed by minimizing the mean square error between image embedding and text embedding, which is expressed as:

[0121]

[0122] Where, represents the distillation loss function; Indicates the sum of N samples in a batch or set, where N is the number of training samples and i is the index of the sample; E image (·) represents an image embedding function or network module; G i is the value corresponding to the i-th input sample x i The sparse gating matrix W(x i ) represents the input image x i The feature map or output tensor produced after the dynamic network; E text (·) represents a text embedding function or a pre-trained text encoder; T i Represents the i-th input image x i The corresponding text description; ||·||2 represents the Euclidean norm, which quantifies the distance between the two embedding vectors in the semantic space. This loss function forces the dynamic network's feature representation to align with the text description in the embedding space, thereby introducing cross-modal semantic supervision and enhancing the network's ability to discriminate fine-grained visual features.

[0123] The cross-modal alignment mechanism constrains the feature learning process of the dynamic network through the semantic information of the text description. For example, when the text description emphasizes a specific part (such as "wheel"), the distillation loss drives the network to enhance the feature response of the corresponding area and i The activation states of related channels are retained. This alignment method can alleviate the problem of semantic information loss caused by sparse computation path selection in dynamic networks, improving the interpretability and generalization ability of the model.

[0124] Preferably, the text encoder maintains frozen parameters during training, optimizing only the feature map layer and sparse gating parameters of the dynamic network to avoid modality conflicts during cross-modal training. The weight initialization of the image embedding map layer can employ an orthogonal initialization strategy to enhance the isotropy of the embedding space, thereby accelerating semantic alignment convergence.

[0125] S6. Dynamically select a calculation path to complete classification reasoning based on the frequency domain energy weight coefficient, sparse gating matrix and tensor decomposition parameters;

[0126] In this embodiment, the selection and classification reasoning process of the dynamic calculation path is implemented by the following technical solutions:

[0127] First, the new frequency domain energy weight coefficient α is calculated in real time based on the input image new , its generation rule is consistent with step S1, that is, by performing fast Fourier transform on the input image and dividing the high-frequency and low-frequency components, the energy distribution ratio is calculated. new Reflects the frequency domain characteristics of the current input image and is used to dynamically control subsequent sparse gating and tensor decomposition parameters.

[0128] Based on α new and learnable channel importance coefficients Generate real-time sparse gating matrix G new ∈{0,1} C×H×W The activation rule of the gating matrix is ​​defined as:

[0129]

[0130] Where, β (n) is the channel importance parameter optimized in step S2, and τ is the dynamically adjusted threshold. The adjustment strategy of the threshold τ is inherited from the activation rate feedback mechanism in step S2, for example, dynamically raising and lowering the threshold according to the preset target sparsity rate to balance computational efficiency and classification accuracy.

[0131] At the same time, combined with the core tensor after asymmetric tensor decomposition and the factor matrix U (1) ,U (2) ,U (3), dynamically reconstruct the convolution kernel weight W decomposed The decomposition rank values ​​R1, R2, and R3 are calculated based on α new Real-time allocation:

[0132]

[0133] Where R max is the preset maximum rank value, constraining the parameter scale after decomposition. Through this allocation rule, the high-frequency dominated image (α new Larger) retains a higher rank value in the output channel dimension to enhance the ability to extract detailed features; low-frequency dominated images compress the output channel rank value to reduce the computational load.

[0134] The reconstructed convolution kernel weight W decomposed The calculation method is:

[0135]

[0136] Where, × n Represents the product operation of the tensor along the nth mode. The reconstructed weights and the real-time gating matrix G new Perform element-wise multiplication operation G new ⊙W decomposed , shielding the weight parameters corresponding to the inactive positions, so that only the key computation paths are retained for reasoning.

[0137] Finally, the filtered weights are applied to the input feature map through a sparse convolution operation, outputting a classification probability distribution. The sparse convolution uses masked multiplication optimization technology to skip the multiplication and addition operations at zero-gated positions, significantly reducing the actual computational effort. Preferably, the sparse computation can utilize a hardware-accelerated instruction set (such as a sparse matrix multiplication and addition unit on a GPU) to achieve real-time inference.

[0138] The dynamic path selection mechanism is based on the frequency domain energy coefficient α new This approach achieves end-to-end adaptation: frequency-domain features drive gating matrix generation and rank assignment, while asymmetric decomposition parameters dynamically adapt to image content characteristics, ultimately maximizing computational efficiency while ensuring classification accuracy. This approach is particularly suitable for edge device deployments, where computing resources must be allocated in real time based on the characteristics of the input data.

[0139] See also Figure 2 The present invention also provides an image big data classification and recognition system based on deep learning, which includes the following modules:

[0140] A frequency domain analysis module, used for performing frequency domain transformation on the input image to generate a frequency domain energy weight coefficient;

[0141] This module converts the input image from the spatial domain to the frequency domain through the fast Fourier transform (FFT) to extract the spectral distribution of high-frequency components and low-frequency components. Specifically, a two-dimensional FFT operation is performed independently on each channel of the input image to generate a frequency domain representation in complex form. Based on the preset frequency threshold range, the central area of ​​the frequency domain coordinate system is defined as the low-frequency component, and the edge area is defined as the high-frequency component, and their energy intensities are calculated separately. The frequency domain energy weight coefficient is generated by the ratio of high-frequency energy to low-frequency energy to quantify the significance ratio of detail features to background features in the image. This coefficient serves as a global control signal for subsequent modules, adapting to image data with different frequency domain characteristics, and is particularly suitable for dynamic resource allocation in edge computing scenarios.

[0142] A dynamic sparse gating module, configured to generate a sparse gating matrix based on the frequency domain energy weight coefficient and the learnable channel importance coefficient to form a gating strategy;

[0143] This module generates a binary, sparse gating matrix based on frequency-domain weight coefficients and learnable channel importance parameters. These channel importance parameters are optimized through backpropagation and characterize the criticality of different feature channels to the classification task. Each element of the gating matrix is ​​activated using a dynamic threshold: if the product of the frequency-domain weight coefficient and the channel importance exceeds the threshold, the corresponding neuronal connection is activated; otherwise, it is blocked. The threshold is adjusted in real time based on the network activation rate. For example, a higher threshold is used in the early stages of training to enhance sparsity, and the threshold is gradually lowered to adapt to the complexity of the task. Through this mechanism, the module can suppress the transfer of redundant features and reduce ineffective computational overhead.

[0144] An asymmetric tensor decomposition module, for performing asymmetric tensor decomposition on convolution kernel weights to generate tensor decomposition parameters, and jointly optimizing them with the gating strategy to construct a dynamic network;

[0145] This module uses dynamic Tucker decomposition to asymmetrically compress convolution kernel weights. Decomposition ranks are dynamically assigned based on frequency-domain weight coefficients: when high frequencies dominate, the output channel dimension retains a higher rank to enhance detail; when low frequencies dominate, the input channel and spatial dimensions are compressed to reduce the number of parameters. The decomposed core tensor and factor matrix are updated via a joint optimization objective function, simultaneously minimizing the weight reconstruction error and the gated sparsity constraint. This design enables the network to adaptively adjust model complexity based on image content characteristics, balancing computational efficiency and feature fidelity.

[0146] a meta-learning task generation module, configured to construct fine-grained meta-learning tasks based on pre-extracted semantic segmentation information and perform task-adaptive training on the dynamic network;

[0147] This module constructs fine-grained meta-learning tasks based on semantic segmentation graphs to improve the model's generalization ability for local features. Specifically, a pre-trained segmentation model is used to extract semantic regions of object parts in the image, and different parts of the same object are defined as independent pseudo-categories. Each meta-task contains a support set and a query set. The support set is used to quickly adapt dynamic network parameters, and the query set is used to verify generalization performance. Using a second-order gradient update algorithm, the gating matrix and decomposition parameters are optimized in the task inner loop, and the global network parameters are updated in the task outer loop, forcing the model to learn a sparse structure generalization strategy across tasks.

[0148] A cross-modal alignment module, configured to semantically align the feature representation of the dynamic network with the cross-modal text description embedding vector to constrain network optimization;

[0149] This module maps image description text into semantic embeddings using a pretrained text encoder and aligns them with the image features output by the dynamic network. After sparse gating, the image features are projected into the same semantic space as the text embeddings through a fully connected layer, constructing a cross-modal distillation loss function. This loss constrains the consistency of image features and text descriptions in the embedding space, introduces semantic supervision from the language modality, and enhances the network's ability to discriminate fine-grained visual features (such as object part attributes).

[0150] The dynamic inference engine module is used to dynamically select the calculation path to complete the classification task according to the frequency domain energy weight coefficient of the input image and the trained gating strategy and tensor decomposition parameters during inference;

[0151] This module is deployed on edge computing devices to achieve low-latency, energy-efficient classification reasoning. Its core components include:

[0152] Lightweight Tensor Decomposition Storage Unit: This unit uses a low-rank format to store dynamically adjusted core tensors and factor matrices, retaining only the parameters corresponding to activation gates, significantly reducing memory usage. Parameter compression and quantization techniques are used to adapt to the storage limitations of edge devices.

[0153] Real-time gating controller: Generates a sparse gating matrix in real time based on the frequency-domain weight coefficients of the input image. Combined with hardware-level sparse computing instruction sets (such as the GPU's mask multiplication unit), it skips multiplication and addition operations at inactive locations, improving computational throughput.

[0154] The dynamic inference engine is deeply optimized for edge computing scenarios, including:

[0155] Computational path compression: Only the decomposed parameters corresponding to the activation gate are retained for inference, reducing data transmission and computational load.

[0156] Parallel pipeline design: Decouple the frequency domain analysis, gate generation, and tensor reconstruction processes into parallel pipeline stages, and utilize multi-core processors to achieve task-level parallelism.

[0157] Energy-efficient adaptive scheduling: Dynamically adjusts the sparsity rate threshold based on the device's real-time power consumption status, further suppressing non-critical computing paths in battery-powered mode to extend device battery life.

[0158] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based image big data classification and recognition method, characterized in that: The method comprises the following steps: S1. Perform frequency domain transformation on the input image, extract the energy distribution of high-frequency components and low-frequency components, and generate frequency domain energy weight coefficients; S2. For the convolution kernel weights in the deep learning network, a sparse gating matrix is ​​dynamically generated based on the frequency domain energy weight coefficients, and a gating strategy is formed by screening activated neuronal connections through learnable channel importance coefficients and adaptive thresholds; S3. Performing asymmetric tensor decomposition on the convolution kernel weights according to the activation state of the sparse gating matrix to obtain tensor decomposition parameters, and synchronously updating the tensor decomposition parameters and the gating strategy by jointly optimizing the objective function to construct a dynamic network. S4. Based on the input image and pre-extracted semantic segmentation information, construct a fine-grained meta-learning task set, and perform task adaptive training on the dynamic network using a second-order gradient update algorithm; S5. Semantically aligning the feature representation of the dynamic network with the pre-generated cross-modal text description embedding vector, and constructing a distillation loss function to constrain and optimize the dynamic network; S6. Dynamically select a calculation path to complete classification reasoning based on the frequency domain energy weight coefficient, sparse gating matrix and tensor decomposition parameters.

2. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S1, the step of generating the frequency domain energy weight coefficient includes: For the input image Perform fast Fourier transform to obtain frequency domain representation Where H, W, and C are the height, width, and number of channels of the image, respectively. The high-frequency components are divided according to the preset frequency range. and low-frequency components The frequency domain energy weight coefficient α is generated by calculating the Frobenius norm ratio of the high-frequency component to the low-frequency component. The expression of the frequency domain energy weight coefficient α is: Where α is the frequency domain energy weight coefficient; ||·|| F represents the Frobenius norm of the matrix; F high is the high frequency component; F low is the low-frequency component; ∈ is a smoothing constant to prevent zero division.

3. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S2, the step of dynamically generating a sparse gating matrix includes: Based on the frequency domain energy weight coefficient α and the learnable channel importance coefficient Generate a sparse gating matrix G∈{0,1} C×H×W , each element G (c,h,w) The activation state of is determined according to the sparse gated element rule, the formula of which is: Where G (c,h,w) Represents the element value at the position with channel index c, height index h, and width index w in the sparse gating matrix G; α is the frequency domain energy weight coefficient; β (c) represents the learnable importance parameter of the cth channel; τ is the dynamically adjusted threshold; The sparse gated element rule combines the frequency domain energy weight coefficient α with the learnable channel importance coefficient Combined, it is determined whether to activate the entire feature channel c by judging whether the product of the two is greater than the dynamically adjusted activation threshold τ; The sparse gated element rule makes the α value larger for images with rich high-frequency details, and the network tends to activate more feature channels for refined analysis; while for images with smooth content, the α value is smaller, and only the most critical channels are retained, thereby suppressing redundant calculations.

4. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S3, the steps of performing asymmetric tensor decomposition on the convolution kernel weights to obtain tensor decomposition parameters, and synchronously updating the tensor decomposition parameters and the gating strategy by jointly optimizing the objective function include: The convolution kernel weight Decomposed into Tucker form: Where W is the convolution kernel weight; the core tensor × n Represents the n-module product of a tensor and a matrix; K represents the spatial dimension of the convolution kernel; K×K means that the height and width of the convolution kernel are both K; the factor matrix C out C represents the number of output channels of the convolutional layer. ln represents the number of input channels of the convolutional layer, and the rank values ​​R1, R2, and R3 are allocated and determined according to the frequency domain energy weight coefficient α according to the rank value allocation rule; wherein, the rank value allocation rule is: Where R max is the preset maximum rank value; Based on the convolution kernel weight W and the objective function, joint optimization is achieved by minimizing the weight reconstruction error with sparsity constraints; wherein the objective function is: Where min W,G Indicates that the objective function is to minimize its value by adjusting W and G, where W is the convolution kernel weight and G represents the sparse gating matrix; ⊙ represents element-by-element multiplication; ||·|| F is the Frobenius norm of the matrix; ||G||0 represents the number of non-zero gates; λ is the sparsity control coefficient.

5. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S4, the step of constructing a fine-grained meta-learning task set includes: Generate a semantic map M∈{0,1} for the input image using pre-trained semantic segmentation information H×W×L , where L represents the number of categories of object parts in the image, h is the height index, and W is the convolution kernel weight; Based on the semantic graph, different parts of the same object are defined as independent categories, and a fine-grained meta-learning task set is constructed. The fine-grained meta-learning task set Include support set and queryset Meta-task set Where S k Represents the kth task T in meta-learning k The support set of (x j ,y j ) represents a sample pair in the support set, x j represents the i-th input sample, y j Represents the input sample x j Corresponding label; y i ,y j ∈{1,2,...,L} is the component category label; N represents the support set S k The number of samples in Q k Represents the kth task T in meta-learning k query set; j is the index of the sample; M represents the query set Q k The number of samples in; T represents the entire meta-learning task set; T k represents the kth meta-task; k is the index of the meta-task; K represents the meta-task.

6. The image big data classification and recognition method based on deep learning according to claim 5 is characterized in that: In step S4, the step of performing task adaptive training on the dynamic network using the second-order gradient update algorithm includes: Based on each meta-task T k , and a second-order gradient update algorithm to update the dynamic network parameters through the inner loop learning rate η, wherein the formula of the second-order gradient update algorithm is: Where θ k Represents the kth meta-task T k The adjusted task-specific model parameters; θ represents the global shared parameters of the model; η is the inner loop learning rate; represents the support set loss function The gradient with respect to the global parameter θ is G is the support set loss function, k is the sparse gating matrix corresponding to the current task; Based on the updated dynamic network parameters and the global parameter update model, the global parameters are updated by the outer loop learning rate γ, wherein the global parameter update model is: Where ← represents the assignment operation; γ is the outer loop learning rate; represents the gradient of the sum of all meta-task query set losses with respect to the global parameter θ, represents the sum of all meta-tasks from the 1st to the Kth, where K represents the total number of meta-tasks. is the loss function of the query set; the gradient calculation includes the second-order derivative term of the sparse gating matrix 7. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S5, the step of semantically aligning the feature representation of the dynamic network with the pre-generated cross-modal text description embedding vector includes: The text description corresponding to the image is converted into a cross-modal text description embedding vector through a pre-trained cross-modal text encoder Where d is the embedding dimension; Embedding vector based on cross-modal text description Mapping the output features of the dynamic sparse network into image embedding vectors Where G i ⊙W(x i ) represents the sparse gating matrix G i ∈{0,1} C×H×W The filtered convolution output; the distillation loss function is constructed by minimizing the mean squared error between the image and text embeddings: Where, represents the distillation loss function; Indicates the sum of N samples in a batch or set, where N is the number of training samples and i is the index of the sample; E image (·) represents an image embedding function or network module; G i is the value corresponding to the i-th input sample x i The sparse gating matrix W(x i ) represents the input image x i The feature map or output tensor produced after the dynamic network; E text (·) represents a text embedding function or a pre-trained text encoder; T i Represents the i-th input image x i Corresponding text description; ||·||2 represents the Euclidean norm; T i is the text description of the i-th sample.

8. The image big data classification and recognition method based on deep learning according to claim 1 is characterized in that: In step S6, the step of dynamically selecting a calculation path according to the frequency domain energy weight coefficient, the sparse gating matrix and the tensor decomposition parameter includes: Calculate the new frequency domain energy weight coefficient α based on the input image new , based on the new frequency domain energy weight coefficient α new and learnable channel importance coefficients Generate real-time sparse gating matrix G new ∈{0,1} C×H×W , wherein the real-time sparse gating matrix G new ∈{0,1} C×H×W The element activation rule in is: when α new β (c) When >τ, G new =1, where τ is the dynamically adjusted threshold; Combined with the core tensor after asymmetric tensor decomposition and the factor matrix U (1) ,U (2) ,U (3) ,pass Reconstruct the convolution kernel weights and perform G new ⊙W decomposed The classification probability distribution is obtained by sparse calculation, where the rank values ​​R1, R2, and R3 are calculated according to Dynamic allocation, R max is the preset maximum rank value, × n Represents the product operation of a tensor along the nth mode.

9. A deep learning-based image big data classification and recognition system, applied to the method according to any one of claims 1 to 8, characterized in that: The system includes the following modules: A frequency domain analysis module, used for performing frequency domain transformation on the input image to generate a frequency domain energy weight coefficient; A dynamic sparse gating module, configured to generate a sparse gating matrix based on the frequency domain energy weight coefficient and the learnable channel importance coefficient to form a gating strategy; An asymmetric tensor decomposition module, for performing asymmetric tensor decomposition on convolution kernel weights to generate tensor decomposition parameters, and jointly optimizing them with the gating strategy to construct a dynamic network; a meta-learning task generation module, configured to construct fine-grained meta-learning tasks based on pre-extracted semantic segmentation information and perform task-adaptive training on the dynamic network; A cross-modal alignment module, configured to semantically align the feature representation of the dynamic network with the cross-modal text description embedding vector to constrain network optimization; The dynamic inference engine module is used to dynamically select the calculation path to complete the classification task according to the frequency domain energy weight coefficient of the input image and the trained gating strategy and tensor decomposition parameters during inference.

10. The deep learning-based image big data classification and recognition system according to claim 9, characterized in that: The dynamic reasoning engine module is deployed on the edge computing device and includes: Lightweight tensor decomposition storage unit, used to store dynamically adjusted core tensors and factor matrices; A real-time gating controller is used to execute the trained gating strategy according to the frequency domain energy weight coefficient of the real-time input image to generate an immediate sparse gating matrix and screen the activated computational path.

Images