Data processing method and system for AI model data

Abstract

The invention relates to the technical field of data processing, in particular to a data processing method and system for AI model data. According to the method, a semantic branch and a quality branch of the asymmetric double-flow heterogeneous network are utilized to respectively generate a prediction confidence coefficient vector and an inherent variance scalar; and performing constraint updating on model parameters through an uncertainty weighted constraint mechanism to realize decoupling of semantic and physical characteristics. The system constructs a two-dimensional evaluation space, and executes local retention, cloud backflow or physical interception instructions according to a data point quadrant through an orthogonal threshold shunt controller controlled by an end-side real-time computing power load. According to the method, high-value cognitive hard cases and low-quality physical noise can be accurately distinguished under low computing power consumption, the effectiveness of cloud training data is effectively improved, and the transmission cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a data processing method and system for AI model data. Background Technology

[0002] In real-world AI model production and cloud platform operation and maintenance scenarios, massive amounts of raw data often contain a lot of noise. If this low-quality data is directly fed into the training process, it will severely affect the model's convergence speed and final accuracy. Therefore, effective cleaning and preprocessing of the data before it enters the annotation platform or training phase is a crucial step.

[0003] Existing data cleaning techniques are mainly divided into two categories: model-feedback-based cleaning and human-interaction-based cleaning. However, they still have significant technical limitations when facing the dual challenges of limited computing power on the edge and high-throughput real-time data access in edge-cloud collaborative scenarios. Some techniques attempt to identify noise by utilizing statistical features after model training, calculating the variance of significant features by improving the neural network structure, and using the magnitude of the variance to measure the model's confidence in the sample, thereby correcting the class label or adjusting the weights. However, the variance calculated by this method reflects the degree of confusion the model has about classifying the current sample. This confusion may stem from physical defects in the data itself, or from the data being a rare sample that the model has never seen before. Existing techniques cannot distinguish between these two situations, leading to high-value, clear, and difficult examples being easily misjudged as dirty data and cleaned, or blurry, unusable data being retained due to forced model fitting, thus reducing the effective information density of the dataset. Other methods employ ensemble learning or cross-validation strategies to improve the accuracy of cleaning, dividing the dataset into multiple subsets and training multiple AI models on the remaining subsets. Only when multiple models consistently predict incorrectly is the data considered dirty data. The significant drawback of this method lies in its extremely high computational and time costs, which are unacceptable in large-scale basic data production scenarios on cloud platforms, and it cannot meet the millisecond-level real-time interception requirements of streaming data access. Some methods introduce model assistance, but still remain dependent on manual intervention. In massive data processing scenarios, this human-machine sequential work mode creates a severe data throughput bottleneck due to manual intervention, preventing true automated data sorting and cleaning.

[0004] In summary, how to distinguish between high-value difficult cases and low-quality physical noise in data with low computing power consumption is a technical problem that urgently needs to be solved in the field of big data cleaning.

[0005] To address this, a data processing method and system for AI model data is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a data processing method and system for AI model data. This invention utilizes the semantic and quality branches of an asymmetric two-stream heterogeneous network to generate a prediction confidence vector and an intrinsic variance scalar, respectively; it updates model parameters through an uncertainty-weighted constraint mechanism, achieving decoupling between semantic and physical features. The system constructs a two-dimensional evaluation space, and through an orthogonal threshold flow controller controlled by the real-time computing power load on the edge, it executes local retention, cloud backflow, or physical interception instructions based on the data point quadrant. This invention can accurately distinguish between high-value cognitive difficulties and low-quality physical noise with low computing power consumption, effectively improving the effectiveness of cloud training data and reducing transmission costs. To achieve the above objectives, the present invention provides the following technical solution: A data processing method for AI model data includes: The raw data is input into a preset semantic-physical feature extractor; the semantic-physical feature extractor includes a main semantic branch and a sub-quality branch. The main semantic branch is used to extract high-dimensional semantic features of the input data to generate a prediction confidence vector, and the sub-quality branch is used to extract low-level texture noise features to generate an intrinsic variance scalar. The parameters of the semantic-physical feature extractor are constrained and updated through an uncertainty weighted constraint mechanism. The uncertainty weighted constraint mechanism uses the inherent variance scalar to perform division decay on the loss term of the main task, and simultaneously uses the regularization term to generate damping constraints on the parameter growth of the sub-quality branch, and iteratively optimizes the inherent variance scalar. The predicted confidence vector and the optimized intrinsic variance scalar are used as inputs to construct a two-dimensional evaluation space, which is then mapped to an orthogonal threshold diversion controller. The orthogonal threshold diversion controller includes orthogonal discrimination rules for dividing data quadrants. The judgment threshold corresponding to the vertical axis is controlled by the real-time computing power load parameters of the end-side device to perform dynamic drift. According to the quadrant where the data point lands in the two-dimensional evaluation space, the corresponding diversion command is executed.

[0007] Preferably, the semantic-physical feature extractor employs an asymmetric two-stream heterogeneous network topology: The main semantic branch employs a deep residual convolutional neural network architecture, extracting high-dimensional semantic representations of the input data through multi-level nonlinear transformations. The network output is connected to a normalized exponential function layer, outputting the predicted confidence vector. The sub-quality branch employs a lightweight shallow convolutional network architecture, with fewer than half the number of layers as the main semantic branch. The branch ends are connected to a global average pooling layer and a non-negative activation function layer, mapping the extracted features to a non-negative intrinsic variance scalar. The main semantic branch and the sub-quality branch share shallow convolutional features at the input, but their parameters are independent in subsequent layers, and they run in parallel during the inference phase.

[0008] Preferably, the secondary quality branch employs a region-of-interest guided frequency-spatial hybrid feature extraction mechanism. The process of generating the intrinsic variance scalar specifically includes: acquiring the class activation map output by the shallow network of the main semantic branch, performing upsampling and binarization processing to generate a spatial attention mask representing the location of the main target; using the spatial attention mask to extract the main region image from the input data and dividing the main region image into several pixel blocks; performing a block-based discrete cosine transform on each pixel block, calculating the ratio of the transform coefficient energy to the DC component energy within the preset coordinate region of the transform matrix, and filtering based on a preset high-frequency noise energy upper limit threshold; simultaneously introducing a standard quantization matrix for... As a priori reference for physical noise, the number of transformation coefficients in the preset coordinate region where the quantized values ​​approach zero is statistically analyzed to generate a frequency domain fingerprint vector representing the physical clarity of the main body region. The original spatial texture feature map of the input data is extracted using a convolutional layer, and the spatial attention mask is multiplied element-wise with the spatial texture feature map to filter background region features and extract weighted texture features containing only the main target. The feature vector obtained by global average pooling of the weighted texture features is concatenated and fused with the frequency domain fingerprint vector in the channel dimension, and the fused features are input into a fully connected layer to regress and generate the intrinsic variance scalar after semantic focusing and frequency band energy calibration.

[0009] Preferably, the uncertainty-weighted constraint mechanism includes the following parameter constraint update process during the training phase: A composite loss function is constructed, including a task prediction error term and a regularization constraint term. The reciprocal of the sum of the intrinsic variance scalar output of the secondary quality branch and a preset smoothing minimum value is used as a dynamic weighting coefficient to weight and attenuate the task prediction error term. This dynamic weighting coefficient is negatively correlated with the intrinsic variance scalar. Simultaneously, a logarithmic term containing the intrinsic variance scalar is introduced into the loss function as a regularization constraint term. The magnitude of this regularization constraint term is monotonically increasing and positively correlated with the intrinsic variance scalar. By minimizing the composite loss function, the weight parameters of the main semantic branch and the secondary quality branch are jointly updated.

[0010] Preferably, the edge device integrates a resource monitoring agent to collect real-time computing load parameters of the edge device at a millisecond frequency, including the remaining battery percentage, NPU computing unit utilization, and current uplink network bandwidth. The orthogonal threshold traffic splitter controller stores a threshold drift curve, defining a negative correlation mapping relationship between the inherent variance judgment threshold and the real-time computing load parameters. When either the remaining battery percentage is lower than a preset warning line or the NPU computing unit utilization is higher than a preset saturation, the controller lowers the inherent variance judgment threshold along the threshold drift curve. When the current uplink network bandwidth is detected to be congested, the controller increases the prediction confidence threshold for the second quadrant data, reducing the total amount of data uploaded to the cloud server.

[0011] Preferably, the orthogonal threshold diversion controller reads the coordinates of data points in the two-dimensional evaluation space in real time and generates control instructions according to a preset quadrant division rule: when a data point falls into the first quadrant consisting of a high prediction confidence interval and a low intrinsic variance interval, a local retention instruction is generated to mark the data as a high-confidence sample and store it in the local inference database of the edge device; when a data point falls into the second quadrant consisting of a low prediction confidence interval and a low intrinsic variance interval, a cloud backflow instruction is generated to mark the data as a cognitive uncertainty sample and establish an encrypted transmission channel with the cloud server through the communication module of the edge device, uploading the data as a high-value difficult example to the cloud's manual annotation interface queue; when a data point falls into the third and fourth quadrants consisting of a high intrinsic variance interval, a physical interception instruction is generated to determine that the data has irreversible physical quality defects and directly release the resources occupied by the data in the edge memory.

[0012] Preferably, the cloud server receives data identified as high-value difficult examples uploaded from multiple end-devices and constructs an incremental training set by combining it with real labels generated by a manual annotation interface; the semantic-physical feature extractor is retrained using the incremental training set, and the network parameters of the main semantic branch and the secondary quality branch are updated through backpropagation algorithm while keeping the uncertainty weighting constraint mechanism unchanged; the updated model parameters are quantized and compressed, and distributed to the end-devices via wireless over-the-air download technology to replace the pre-installed semantic-physical feature extractor in the end-devices.

[0013] Preferably, a data processing system for AI model data includes: Feature extraction module: Input the raw data into a preset semantic-physical feature extractor; the semantic-physical feature extractor includes a main semantic branch and a sub-quality branch. The main semantic branch is used to extract high-dimensional semantic features of the input data and generate a prediction confidence vector. The sub-quality branch is used to extract low-level texture noise features and generate an intrinsic variance scalar. Parameter optimization module: The parameters of the semantic-physical feature extractor are updated by means of an uncertainty weighted constraint mechanism. The uncertainty weighted constraint mechanism uses the inherent variance scalar to perform division decay on the loss term of the main task, and simultaneously uses the regularization term to generate damping constraints on the parameter growth of the sub-quality branch, and iteratively optimizes the inherent variance scalar. The traffic splitting control module constructs a two-dimensional evaluation space using the predicted confidence vector and the optimized intrinsic variance scalar as inputs, and maps it to an orthogonal threshold traffic splitting controller. The orthogonal threshold traffic splitting controller includes orthogonal discrimination rules for dividing data quadrants, where the decision threshold corresponding to the vertical axis is controlled by the real-time computing power load parameters of the edge device. Based on the quadrant where the data points fall in the two-dimensional evaluation space, the module executes corresponding traffic splitting instructions. Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an asymmetric two-stream network structure comprising a main semantic branch and a secondary quality branch, and combines this with a composite loss function based on an uncertainty-weighted constraint mechanism, endowing the model with the ability to simultaneously perceive semantic content and physical quality in a single inference iteration. This invention utilizes the secondary quality branch to directly regress the inherent variance scalar of the predicted data, enabling accurate identification of inherent defects such as motion blur, occlusion, or sensor noise at the physical signal level. This mechanism allows the system to orthogonally separate clear but unseen hard examples from blurred, unusable images in a two-dimensional evaluation space, thereby cleaning up low-quality data while maximizing the retention of hard example data highly valuable for model iteration, significantly improving the effective information density of the training dataset.

[0014] 2. This invention abandons the computationally expensive cleaning path required by existing technologies, which necessitates repeated training of multiple models or K-fold cross-validation. Instead, it innovatively employs a single forward propagation to complete quality assessment. In particular, the secondary quality branch utilizes a lightweight, shallow architecture with fewer than half the layers of the main branch, significantly reducing the number of model parameters and computational load. This design allows the method to be easily deployed on edge devices with limited computing power, such as smart cameras and vehicle-mounted units, achieving millisecond-level real-time data interception. Compared to traditional methods, this invention reduces the computational cost of the cleaning process by several orders of magnitude, prevents junk data from leaving the edge, and significantly saves on edge-to-cloud transmission bandwidth and cloud storage costs.

[0015] 3. This invention utilizes an orthogonal threshold traffic splitting controller to establish an unattended decision-making logic that automatically executes local retention, cloud backflow, and physical interception based on the data point quadrant. This controller not only automatically splits traffic based on quality scores but also dynamically determines thresholds based on the real-time computing load of edge devices, achieving adaptive adjustment to environmental changes. This mechanism completely replaces the inefficient manual review and correction steps in existing technologies, significantly improving data processing throughput and ensuring the continuity and stability of core business operations even under harsh conditions or resource-constrained scenarios, forming an efficient edge-cloud collaborative closed loop. Attached Figure Description

[0016] Figure 1 A flowchart of a data processing method for AI model data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a data processing system for AI model data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an orthogonal threshold splitting strategy provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figures 1 to 3 This invention provides a data processing method for AI model data, the technical solution of which is as follows: A data processing method for AI model data includes: The raw data is input into a preset semantic-physical feature extractor; the semantic-physical feature extractor includes a main semantic branch and a sub-quality branch. The main semantic branch is used to extract high-dimensional semantic features of the input data to generate a prediction confidence vector, and the sub-quality branch is used to extract low-level texture noise features to generate an intrinsic variance scalar. The parameters of the semantic-physical feature extractor are constrained and updated through an uncertainty weighted constraint mechanism. The uncertainty weighted constraint mechanism uses the inherent variance scalar to perform division decay on the loss term of the main task, and simultaneously uses the regularization term to generate damping constraints on the parameter growth of the sub-quality branch, and iteratively optimizes the inherent variance scalar. The predicted confidence vector and the optimized intrinsic variance scalar are used as inputs to construct a two-dimensional evaluation space, which is then mapped to an orthogonal threshold diversion controller. The orthogonal threshold diversion controller includes orthogonal discrimination rules for dividing data quadrants. The judgment threshold corresponding to the vertical axis is controlled by the real-time computing power load parameters of the end-side device to perform dynamic drift. According to the quadrant where the data point lands in the two-dimensional evaluation space, the corresponding diversion command is executed.

[0019] Example 1

[0020] This invention provides an AI model data processing method based on edge-cloud collaboration. This method is widely used in scenarios with high real-time requirements and limited bandwidth resources, such as intelligent traffic monitoring, autonomous driving data acquisition, and industrial quality inspection. The following detailed explanation uses an intelligent traffic monitoring system and its backend AI model production platform as examples.

[0021] As one embodiment of the present invention, refer to Figure 1 A flowchart of a data processing method for AI model data, see reference. Figure 2 A schematic diagram of the structure of a data processing system for AI model data, referring to... Figure 3 A schematic diagram of an orthogonal threshold splitting strategy.

[0022] The front-end equipment consists of embedded smart cameras (edge ​​devices) deployed on highway gantry frames, while the back-end consists of AI training clusters (cloud servers) deployed in data centers. The core problem to be solved is: how to avoid wasting transmission bandwidth and annotation costs on low-quality images in rainy or foggy weather, when high-speed motion is blurry, or when the camera is damaged, while accurately selecting clear but difficult-to-identify "high-value difficult examples" for model iteration.

[0023] The “difficult examples” (or “high-value cognitive difficult examples” or “cognitive uncertainty samples”) mentioned in this specification specifically refer to data samples that are clear and complete in terms of physical imaging quality, but whose model cannot make high-confidence classification predictions due to insufficient feature extraction capabilities of the model itself or bias in the distribution of training data.

[0024] Furthermore, the semantic-physical feature extractor employs an asymmetric two-stream heterogeneous network topology: The main semantic branch employs a deep residual convolutional neural network architecture, extracting high-dimensional semantic representations of the input data through multi-level nonlinear transformations. The network output is connected to a normalized exponential function layer, outputting the predicted confidence vector. The sub-quality branch employs a lightweight shallow convolutional network architecture, with fewer than half the number of layers as the main semantic branch. The branch ends are connected to a global average pooling layer and a non-negative activation function layer, mapping the extracted features to a non-negative intrinsic variance scalar. The main semantic branch and the sub-quality branch share shallow convolutional features at the input, but their parameters are independent in subsequent layers, and they run in parallel during the inference phase.

[0025] The main semantic branch employs a ResNet-50 deep residual convolutional neural network architecture, containing 49 convolutional layers and residual skip connections, with a total of approximately 25.5 million parameters. The network input receives an RGB image with a resolution of H×W×3. The shallow convolutional layers following the first residual block (stride=1) output feature dimensions of... (in =64, = =1 / 4 of the original resolution). The deep layers of the main semantic branch (the second to fourth Bottleneck groups) independently extract high-dimensional semantic features. The network output is connected to a global average pooling and Softmax normalized exponential function layer to generate a prediction confidence vector with N classes (e.g., 80 dimensions).

[0026] The sub-quality branch employs a 4-layer convolutional network architecture, with kernel sizes of [3×3, 3×3, 3×3, 1×1] and channel numbers of [64, 128, 256, 1], totaling approximately 1.1 million parameters. The sub-quality branch shares the output features (64×10 ... × After this, the two branches perform convolution operations in parallel and independently in space. The global average pooling layer at the end of the sub-quality branch compresses the 2D feature map into a scalar, followed by a ReLU activation function to ensure that the output is non-negative, generating a normalized intrinsic variance scalar in the range [0, 1].

[0027] Specifically, the extractor is deployed on an edge AI camera chip with limited computing resources, processing real-time video streams with a resolution of 1920×1080. The main semantic branch is configured as a ResNet-50 deep residual network with approximately 25.5 million parameters, used for fine-grained identification of dozens of semantic categories, such as hazardous chemical vehicles, cars, and pedestrians. For example, when a special engineering vehicle with an unusual appearance drives by at high speed, the main semantic branch, after deep feature extraction, outputs a prediction confidence vector with a maximum value of only 0.42, indicating that the model's semantic understanding of the target has high uncertainty.

[0028] Meanwhile, the sub-quality branch is designed as an ultra-lightweight network with only 5 convolutional layers, containing less than 5% of the parameters of the main branch and extremely low computational cost. This branch runs in parallel, focusing on detecting physical defects such as motion blur and low-light noise. For example, although the aforementioned engineering vehicle image is semantically difficult to understand, its image quality is clear and the edges are sharp. The intrinsic variance scalar output by the sub-quality branch after inference is 0.05 (a normalized value, the lower the value, the better the quality). Conversely, if the image is blurred due to windshield wipers obstructing the view at night during heavy rain, the sub-branch will output an intrinsic variance scalar as high as 0.88. Utilizing this asymmetric structure, while keeping the total inference time per frame within 35 milliseconds, by comparing two drastically different data combinations, "0.42 confidence + 0.05 variance" and "0.42 confidence + 0.88 variance," the former is accurately identified as a high-value difficult example that needs to be uploaded to the cloud, while the latter is a physical waste image that should be discarded directly. This achieves efficient data cleaning and bandwidth saving at the edge.

[0029] This invention employs an asymmetric network topology, specifically a deep main branch and a shallow sub-branch, which significantly optimizes the inference efficiency and computing power allocation of edge devices. Since physical quality assessment primarily relies on low-level texture features and does not require deep abstraction like semantic recognition, a lightweight shallow network suffices. Compared to traditional two-stream peer-to-peer networks, this design significantly reduces the number of model parameters and FLOPs, enabling this complex data cleaning algorithm to be deployed on computationally limited embedded chips or mobile terminals, meeting real-time processing requirements.

[0030] Furthermore, the secondary quality branch employs a region-of-interest guided frequency-spatial hybrid feature extraction mechanism. The process of generating the intrinsic variance scalar specifically includes: acquiring the class activation map output by the shallow network of the main semantic branch, performing upsampling and binarization processing to generate a spatial attention mask representing the location of the main target; using the spatial attention mask to extract the main region image from the input data and dividing the main region image into several pixel blocks; performing a block-based discrete cosine transform on each pixel block, calculating the ratio of the transform coefficient energy to the DC component energy within the preset coordinate region of the transform matrix, and filtering it in conjunction with a preset high-frequency noise energy upper limit threshold; simultaneously introducing a standard quantization matrix as... As a priori reference for physical noise, the number of transformation coefficients in the preset coordinate region where the quantized values ​​approach zero is statistically analyzed to generate a frequency domain fingerprint vector representing the physical clarity of the main body region. The original spatial texture feature map of the input data is extracted using a convolutional layer, and the spatial attention mask is multiplied element-wise with the spatial texture feature map to filter background region features and extract weighted texture features containing only the main target. The feature vector obtained by global average pooling of the weighted texture features is concatenated and fused with the frequency domain fingerprint vector in the channel dimension, and the fused features are input into a fully connected layer to regress and generate the intrinsic variance scalar after semantic focusing and frequency band energy calibration.

[0031] The "intrinsic variance scalar" in this invention is actually a comprehensive "physical quality score," whose value reflects the degree of image sharpness defect after comprehensive analysis of frequency and spatial domain features. This scalar's value is not a statistically significant "variance," but rather a scoring index within the range of [0, 1]. Its calculation is based on the following physical principles: Frequency domain principle: According to signal processing theory, image sharpness is positively correlated with the energy density of its high-frequency components. Physical defects such as motion blur, low-resolution input, and sensor Gaussian noise can all lead to the loss of high-frequency components. Spatial domain principle: By limiting the analysis range to the main target area through a spatial attention mask, interference from legitimate low-frequency background features (such as snow and blue sky) is eliminated, making the score more focused on the physical quality of the subject. Comprehensive evaluation: By fusing frequency domain fingerprints (high-frequency energy percentage, zero-touch rate, etc.) with spatial domain texture features (edge ​​sharpness, texture contrast), a multi-dimensional quality assessment is formed, improving the accuracy to approximately 92% (based on measured data).

[0032] Therefore, although based on frequency domain analysis, the comprehensive introduction of spatial constraints and multidimensional fusion significantly reduces the misclassification rate of legitimate low-frequency images. The scalar is named "intrinsic variance" to correspond with "prediction confidence" (both are single scalar outputs), but its physical meaning should be more accurately understood as "physical quality score" or "sharpness index".

[0033] Specifically, firstly, frequency domain feature extraction is performed: the target region in the input image, determined by the attention mask, is divided into several 8×8 pixel blocks, and a two-dimensional discrete cosine transform is performed on each pixel block. For the transformed coefficient matrix, the triangular region below the sub-diagonal of the lower right corner of the matrix is ​​defined as the high-frequency coordinate region. The sum of the squares of all transform coefficients in this region is calculated as the high-frequency energy, and the square of the DC component in the upper left corner of the matrix is ​​calculated as the DC energy. The ratio of the high-frequency energy to the sum of the DC energy and the preset smoothing minimum value is calculated. If this ratio is greater than the preset upper limit threshold of high-frequency noise energy (e.g., 0.15), the pixel block is marked as a high-frequency anomalous block and excluded from subsequent statistics. Next, a standard JPEG luminance quantization matrix is ​​introduced to perform division and rounding quantization operations on the transform coefficients of the retained pixel blocks. The proportion of coefficients whose quantized values ​​become zero in the high-frequency coordinate region is counted to the total number of coefficients in the region, generating a frequency domain fingerprint vector characterizing the physical sharpness of the target region. Next, spatial feature extraction is performed: the original spatial texture feature map of the input data is extracted using a convolutional layer, and the spatial attention mask is multiplied element-wise with the spatial texture feature map to filter background region features. After global average pooling, a spatial feature vector is obtained. Finally, the frequency domain fingerprint vector and the spatial feature vector are concatenated and fused along the channel dimension. The fused feature vector is input into a fully connected layer containing multiple neurons for feature mapping, and the output of the fully connected layer is input into a sigmoid activation function. The output value is constrained to an open interval between 0 and 1 to obtain the intrinsic variance scalar. This scalar measures the probability that the data contains physical noise defects; the closer the value is to 1, the more severe the defect.

[0034] The preset coordinate region specifically refers to the lower right corner region of the DCT transform coefficient matrix, which represents mid-to-high frequency information. Taking an 8×8 transform matrix as an example, this region is defined as a triangular area extending from the secondary diagonal of the matrix (the diagonal connecting the lower left corner to the upper right corner) to the lower right vertex of the matrix. The coefficients in this region reflect the edge sharpness and detail texture of the image, but do not include the DC component and low-frequency contour component located in the upper left corner of the matrix.

[0035] For each pixel block, a block-based discrete cosine transform (DCT) is performed, using an 8×8 pixel block size to generate 64 DCT transform coefficient matrices. A preset coordinate region is defined as a triangular area extending downwards and to the right from the secondary diagonal (connecting positions (0, 7) to (7, 0)) within the 8×8 matrix. Specifically, this region includes all coefficient positions from (1, 7), (2, 6), (2, 7)... to (7, 7), totaling approximately 28 positions. This region represents the mid-to-high frequency information of the image, and its energy value reflects the edge sharpness and detail texture of the image.

[0036] Specifically, in the specific application scenario of intelligent traffic monitoring, this embodiment provides a detailed description of a typical working condition: nighttime highway vehicle capture. It is assumed that the edge device is an intelligent capture unit deployed on a gantry, equipped with a computationally limited embedded NPU. When the device acquires a vehicle driving image with a resolution of 1920×1080, the main semantic branch first performs lightweight inference on the image, outputting a class activation map with a resolution of 7×7. This activation map is then upsampled to the original image size, and a threshold of 0.5 is set for binarization processing to generate a spatial attention mask that only covers the main body of the vehicle (including license plate, headlights, and vehicle outline), thereby removing the black road surface and night sky background that occupy a large proportion of the image.

[0037] Subsequently, the sub-quality branch uses this mask to extract the effective pixel region of the vehicle body and divides it into thousands of 8×8 pixel blocks. The NPU calls the underlying image encoding and decoding instruction set to perform block discrete cosine transform (DCT) in parallel on each pixel block. The triangular region in the lower right corner of the preset transform matrix is ​​the high-frequency coordinate region, which corresponds to the edge sharpness and detail texture of the image. In this calculation process, the ratio of the transform coefficient energy to the DC component energy within the preset coordinate region of the transform matrix is ​​calculated and filtered in conjunction with a preset high-frequency noise energy upper limit threshold. If the high-frequency energy of some pixel blocks is detected to be abnormally high and exceeds the threshold, they are excluded or truncated to prevent high-frequency random noise from being misjudged as valid high-resolution features.

[0038] Next, when processing a "discarded" image with severe motion blur due to excessive vehicle speed, after the aforementioned filtering, the effective high-frequency energy was found to be extremely low, with a ratio to the DC component below 0.05. At this point, a standard JPEG luminance quantization table was introduced as a physical noise reference prior, and the transform coefficients were divided by the corresponding step size in the quantization table. Since the original high-frequency coefficients were extremely small, after performing division and rounding operations, the number of transform coefficients in the preset coordinate region where the quantized values ​​approached zero was statistically analyzed. The results showed that over 95% of the pixel blocks had coefficients of zero in the high-frequency region. This statistical result was encoded as a frequency domain fingerprint vector, clearly indicating a severe loss of physical sharpness in the main subject area.

[0039] Meanwhile, the spatial texture features extracted by the convolutional layer, after mask filtering, retain only the smooth, blurred texture of the vehicle body area, eliminating random noise interference from the background caused by high sensitivity. Finally, the frequency domain fingerprint and spatial features are fused in a fully connected layer, resulting in a regression output with an intrinsic variance scalar as high as 0.92. This value far exceeds the set physical interception threshold of 0.8. The orthogonal threshold splitter controller then determines that the image is data with physical quality defects and directly performs a memory release operation on the edge, avoiding invalid data uploads to the cloud and effectively saving approximately 4MB of transmission bandwidth and cloud storage space.

[0040] This invention significantly improves the physical interpretability and anti-interference capability of data cleaning by employing a region-of-interest-guided frequency-spatial hybrid feature extraction mechanism. First, by utilizing a spatial attention mask, frequency domain analysis is strictly limited to the main target area, effectively solving the technical problem of traditional global frequency domain analysis easily misjudging background blurring or large areas of solid-color background in photography as high-frequency loss, ensuring that the cleaning strategy focuses on core business objects. Second, by using block-based discrete cosine transform instead of computationally intensive Fourier transform and introducing a standard quantization matrix as a priori reference, this not only fully utilizes the hardware acceleration capabilities of the edge chip, significantly reducing inference latency and energy consumption, but also enables the model to accurately quantize motion blur and moiré noise from the underlying logic of signal processing, achieving precise orthogonal decoupling between semantically difficult examples and physically unusable data.

[0041] Furthermore, the uncertainty-weighted constraint mechanism's parameter constraint update process during the training phase includes: A composite loss function is constructed, including a task prediction error term and a regularization constraint term. The inverse of twice the sum of the intrinsic variance scalar output of the secondary quality branch and a preset smoothing minimum is calculated as a dynamic weighting coefficient. This dynamic weighting coefficient is used to perform a multiplication operation on the main task loss term to achieve attenuation, i.e., dividing the main task loss term by twice the sum of the intrinsic variance scalar and the preset smoothing minimum to obtain a weighted task error term. A composite loss function is constructed, which is composed of the weighted task error term, the regularization constraint term, and the orthogonal projection constraint term. The regularization constraint term is the product of the natural logarithm of the sum of the intrinsic variance scalar and the preset smoothing minimum and a preset regularization coefficient. By minimizing the composite loss function, the weight parameters of the main semantic branch and the secondary quality branch are jointly updated.

[0042] Specifically, the calculation logic of the composite loss function follows the principle of uncertainty learning, and it consists of the sum of a task error weighting term and a regularization penalty term. The task error weighting term is calculated by dividing the task prediction error of the main semantic branch by twice the sum of the intrinsic variance scalar of the secondary quality branch output and a preset smoothing constant. The regularization penalty term is calculated by taking half the natural logarithm of the intrinsic variance scalar. Finally, the model updates its parameters by minimizing the sum of these two terms. In this calculation process, the smoothing constant in the denominator is used to prevent numerical overflow caused by the variance approaching zero.

[0043] When training on a massive dataset containing ten million traffic scene images, this uncertainty-weighted constraint mechanism acts as an intelligent gradient filter. Suppose that in the 50,000th iteration, the model encounters a dirty data sample with severe water fog obscuring the image due to heavy rain. At this point, the main semantic branch cannot extract effective vehicle body features, causing the prediction to deviate significantly from the true label, resulting in a raw task prediction error as high as 3.5. In traditional training, such a large error would generate a strong backpropagation gradient, forcing the model to incorrectly memorize the water fog texture as a classification criterion.

[0044] However, in this scheme, the secondary quality branch keenly detects the lack of high-frequency information in the image, outputting a judgment that the image has significant physical uncertainty, pushing the intrinsic variance scalar up to 0.92. The reciprocal of the sum of this scalar and a preset smoothing minimum (e.g., 1e-6) is immediately used as a dynamic weighting coefficient to strongly attenuate the aforementioned error of 3.5. Because the denominator becomes larger, the weighted error participating in the final weight update is reduced to an extremely low level, effectively ignoring this erroneous prediction, protecting the model parameters from noise contamination, and also avoiding the risk of numerical overflow due to some extreme sample variances approaching zero.

[0045] Conversely, when the model encounters a high-value hard example, such as a rare car model with an intact exterior but key parts obscured by a lamppost, the main branch may still produce a prediction error as high as 3.5 due to incomplete semantic features. In this case, to reduce the current loss function value, the model might "attempt to cheat," that is, tend to output a large intrinsic variance scalar, attempting to mathematically attenuate this large error by increasing the denominator. However, the logarithmic regularization term containing the intrinsic variance scalar in the composite loss function immediately comes into play. Since the image is sharp and clear in texture, the physical features extracted by the sub-branch do not support a high variance output. Once the model forcibly increases the variance scalar from 0.1 to 0.9, the value of this logarithmic term rises sharply, causing the total loss to increase instead of decrease. Guided by the gradient, the model is forced to accept that the image is sharp, maintaining the intrinsic variance scalar at a low level of around 0.15. This results in the 3.5 prediction error not only not being attenuated, but being amplified by the highly weighted reciprocal coefficient (approximately 6.6 times), generating a huge gradient signal. This signal forces the main semantic branch to significantly adjust its deep parameters to learn the vehicle body inference logic under occlusion conditions. Through this differentiated constraint update, after 100 epochs of training, the model maintains its robustness to rain, fog, and noise while improving the accuracy of identifying difficult occlusion cases by 18.5% compared to the baseline model.

[0046] The regularization constraint term also includes an orthogonal projection constraint mechanism. When constructing the composite loss function, the cosine similarity between the feature vector output by the main semantic branch and the feature vector output by the secondary quality branch is calculated, and the absolute value of this similarity is added to the composite loss function as a penalty term. By minimizing this penalty term, the main semantic features and the secondary quality features are forced to maintain an orthogonal relationship in the high-dimensional vector space, so that the main semantic branch focuses only on semantic category features and the secondary quality branch focuses only on physical texture features that are not semantically related, thereby achieving decoupling of the two types of features.

[0047] Specifically, to prevent "feature entanglement" during model training—that is, to prevent the secondary quality branch from "guessing" quality using semantic features, or the main semantic branch from using background texture to assist classification—this embodiment introduces orthogonal projection constraints during the training phase. The semantic feature vector from the penultimate layer of the main semantic branch and the quality feature vector from the end of the secondary quality branch are extracted in real time, and the cosine of the angle between these two vectors in the feature space is calculated. If the two vectors tend to be in the same or opposite directions, it indicates that the information extracted by both is highly overlapping, and the orthogonal penalty term in the loss function will increase significantly. During backpropagation, this penalty forces the network to adjust its parameters until the two feature vectors are spatially perpendicular. For example, for a sample of a "vehicle in the rain," through orthogonal constraints, the main branch is forced to extract only identity-related semantic information such as the "vehicle outline," while completely stripping the signal generated by "raindrop occlusion" to the secondary branch for processing. This ensures that the intrinsic variance scalar purely reflects physical quality, unaffected by vehicle type.

[0048] This invention further introduces an orthogonal projection constraint mechanism, mathematically forcing a complete decoupling of semantic and physical quality information from the data. This effectively solves the problem that the model might exploit statistical patterns in background texture to cheat and reduce the loss function, ensuring that the intrinsic variance scalar output of the sub-quality branch has high independence and physical interpretability, thereby improving the clarity and accuracy of subsequent quadrant partitioning in the two-dimensional evaluation space.

[0049] This invention designs a composite loss function including division decay and logarithmic regularization, endowing the model with noise-resistant adaptive capabilities during training and preventing overfitting to dirty data. By dynamically adjusting the weights of the main task loss using the intrinsic variance scalar, it automatically reduces the contribution of unrecognizable physically noisy images to the gradient, equivalent to soft dropout. Simultaneously, the regularization constraint prevents the model from cheating by blindly outputting high variance to evade penalties. This balancing mechanism ensures that the model can focus on learning samples that are clear in image but semantically difficult to distinguish, improving the model's final generalization accuracy.

[0050] Furthermore, the edge device integrates a resource monitoring agent to collect real-time computing load parameters of the edge device at millisecond-level frequency, including the remaining battery percentage, NPU computing unit utilization, and current uplink network bandwidth. The orthogonal threshold traffic splitter controller stores a threshold drift curve, defining a negative correlation mapping relationship between the inherent variance judgment threshold and the real-time computing load parameters. When either the remaining battery percentage is lower than a preset warning line or the NPU computing unit utilization is higher than a preset saturation level, the controller lowers the inherent variance judgment threshold along the threshold drift curve. When the current uplink network bandwidth is detected to be congested, the controller increases the prediction confidence threshold for the second quadrant data, reducing the total amount of data uploaded to the cloud server.

[0051] The threshold drift curve is specifically as follows: when the remaining battery percentage is detected to be greater than or equal to a first power threshold (e.g., 30%), the inherent variance determination threshold is maintained at a first preset value (e.g., 0.85); when the remaining battery percentage is detected to be between a second power threshold and the first power threshold (e.g., in the range of 20% to 30%), the inherent variance determination threshold is linearly lowered to a second preset value (e.g., 0.55); when the remaining battery percentage is detected to be lower than the second power threshold (e.g., lower than 20%), the inherent variance determination threshold is further lowered to a third preset value (e.g., 0.25); wherein the first preset value is greater than the second preset value, and the second preset value is greater than the third preset value, thereby achieving a negative correlation adjustment in which the inherent variance determination threshold decreases in a step-like or linear manner as the battery level decreases.

[0052] Specifically, in mobile electronic police cameras deployed in the field using solar power, the resource monitoring agent is configured to poll the status every 50 milliseconds. Assuming a morning with continuous rain leading to insufficient charging, the device detects that the remaining battery power has dropped to 15% (below the preset warning line of 20%), while the NPU utilization surges to 92% due to heavy traffic during rush hour. At this point, the orthogonal threshold shunt controller immediately triggers the threshold drift mechanism, lowering the inherent variance judgment threshold from the default 0.85 to 0.40 based on the internally stored energy efficiency priority curve. This adjustment means entering a strict screening mode; even slightly jittery images (variance 0.5) will be judged as third-quadrant data and directly intercepted, thereby significantly reducing subsequent processing energy consumption and ensuring the device does not shut down due to depleted power.

[0053] Furthermore, when encountering peak-hour network congestion at base stations, and the available uplink bandwidth drops sharply from 10Mbps to 300Kbps, the controller automatically adjusts the admission logic in the second quadrant. The lower limit for prediction confidence is increased from 0.05 to 0.35. This strategy is based on the logic that, when bandwidth is extremely limited, extremely low-quality difficult examples where the model is almost entirely guessing (confidence 0.05-0.35) are discarded, while only high-value "partially understood" samples where the model has some basis but is not yet certain (confidence 0.35-0.60) are retained. Real-world testing shows that this dynamic strategy successfully reduced the total number of backhaul data packets per unit time by approximately 75% without affecting the accuracy of the core model iteration, effectively preventing congestion and packet loss in the communication link.

[0054] The orthogonal threshold split controller also performs baseline drift calibration based on ambient noise: the resource monitoring agent maintains a sliding time window, statistically analyzes the distribution of the intrinsic variance scalar of all input data within the window, and calculates the noise baseline value of the distribution; the vertical axis decision threshold in the orthogonal discrimination rule is composed of the basic threshold determined by the real-time computing power load parameters and the noise baseline value; when the ambient lighting conditions deteriorate, causing the overall intrinsic variance distribution of the input data to shift to the right, the controller automatically adjusts the vertical axis decision threshold to counteract the interference of ambient noise on data quality judgment.

[0055] Maintain a sliding time window of size W = 3600 seconds (1 hour), and define the noise baseline value as baseline = Q1 (lower quartile). Check the baseline for significant changes every 5 minutes. If the baseline increases by more than 20%, the environmental noise is considered to have significantly deteriorated.

[0056] Specifically, considering the drastic changes in outdoor ambient light, an environmental adaptive calibration mechanism was introduced. For example, in late night or foggy weather, due to the physical characteristics of the image sensor, even clear images will have a lower overall signal-to-noise ratio, resulting in a generally higher intrinsic variance scalar output from the sub-quality branch (e.g., drifting from 0.05 during the day to 0.3). If a fixed threshold (e.g., 0.2) is used, these normal nighttime samples will be incorrectly intercepted as physical defects. To address this, a sliding window with a duration of 1 hour is maintained to statistically analyze the distribution histogram of intrinsic variance in real time. When a significant increase in the lower quartile of the variance of all samples in the most recent hour is detected, it is determined that the environmental noise floor has increased. At this time, the controller will add a dynamic bias to the original threshold based on computing power load, automatically raising the interception threshold, thereby ensuring that relatively clear difficult examples can still be filtered out even in harsh environments, rather than simply and crudely intercepting all data.

[0057] This invention proposes an environment-adaptive baseline drift calibration technique, enabling edge devices to sense changes in the physical environment. By dynamically adjusting the baseline based on the distribution patterns of historical data, it eliminates background noise interference caused by environmental factors such as day-night cycles and weather changes. This avoids the erroneous decision to discard all normal data due to a decrease in the global signal-to-noise ratio under low light or severe weather conditions, significantly improving robustness and business continuity under all-weather, all-time operating conditions.

[0058] This invention introduces a resource monitoring agent and threshold drift curve, deeply binding the algorithm strategy with the physical state of the hardware, achieving high availability and environmental adaptability. In IoT scenarios, hardware resources fluctuate dynamically. When power is critically low or the network is congested, a fixed threshold can lead to task backlog or device downtime. This solution dynamically lowers the threshold or tightens the upload strategy, automatically degrading operation when resources are limited, prioritizing the uninterrupted operation of core services, and minimizing the power consumption of transmitting unnecessary data, thus extending the battery life of the edge devices.

[0059] Before generating the cloud backflow instruction, a deduplication step based on perceptual hashing is performed: the orthogonal threshold splitting controller extracts the perceptual hash fingerprint of the region of interest of the current data point and calculates the Hamming distance with the hash fingerprint of the uploaded data in the cache queue of the edge device; if the Hamming distance is less than the preset similarity threshold, the current data is determined to be a redundant sample, the controller intercepts the cloud backflow instruction and only updates the local statistical count; if the Hamming distance is greater than the preset similarity threshold, the upload operation is performed and the current perceptual hash fingerprint is updated to the cache queue.

[0060] The desampling logic is as follows: A 64-bit hash fingerprint is generated using a differential hash algorithm. A circular cache queue of size N (e.g., 50) is maintained. For the current data, the Hamming distance between it and all fingerprints in the cache queue is calculated. If the minimum Hamming distance is less than a preset similarity threshold T (e.g., 8 bits, representing 87.5% similarity), the current data is determined to be a redundant sample.

[0061] Specifically, in traffic congestion or at traffic light intersections, vehicles may remain within the camera's field of view for extended periods. To avoid continuously uploading hundreds of duplicate images of the same difficult vehicle, wasting valuable uplink bandwidth, the controller incorporates a desampling module. When an image is identified as a high-value difficult example in the second quadrant, the main vehicle region in the image is first down-sampled and subjected to discrete cosine transform to generate a perceptual hash fingerprint in character form. This fingerprint is then compared with the 50 most recent fingerprints uploaded within the past 5 minutes to calculate the Hamming distance. If the current vehicle's fingerprint differs from the fingerprint uploaded in the previous frame by only 2 characters (Hamming distance of 2, less than the preset threshold of 5), the image is considered visually repetitive, even though it is a difficult example, it is a redundant one. In this case, only one difficult example occurrence event is recorded in the log, and image data is not sent over the network until the vehicle moves significantly or the viewing angle changes, causing the Hamming distance to exceed the threshold, at which point uploading is triggered again.

[0062] This invention adds a perceptual hash-based desampling strategy to the cloud backhaul mechanism, effectively solving the problem of repeatedly uploading large amounts of highly similar data for the same target in low-speed or static scenarios. This not only further saves communication traffic and power consumption of edge devices, but also prevents the cloud training set from being overwhelmed by a large number of duplicate samples, ensuring the diversity and effectiveness of the backhaul dataset, and maximizing the utilization efficiency of cloud annotation and training resources.

[0063] Furthermore, the orthogonal threshold diversion controller reads the coordinates of the data points in the two-dimensional evaluation space in real time and generates control instructions according to the preset quadrant division rules: when the data point falls into the first quadrant consisting of a high prediction confidence interval and a low intrinsic variance interval, a local retention instruction is generated to mark the data as a high-confidence sample and store it in the local inference database of the edge device; when the data point falls into the second quadrant consisting of a low prediction confidence interval and a low intrinsic variance interval, a cloud backflow instruction is generated to mark the data as a cognitive uncertainty sample and establish an encrypted transmission channel with the cloud server through the communication module of the edge device, uploading the data as a high-value difficult example to the cloud's manual annotation interface queue; when the data point falls into the third and fourth quadrants consisting of a high intrinsic variance interval, a physical interception instruction is generated to determine that the data has irreversible physical quality defects and directly release the resources occupied by the data in the edge memory.

[0064] The two-dimensional evaluation space uses the maximum value of the predicted confidence vector as the horizontal axis (X-axis) and the intrinsic variance scalar as the vertical axis (Y-axis). The lower limit of the dynamically drifted predicted confidence is used as the horizontal axis dividing line, and the intrinsic variance threshold is used as the vertical axis dividing line, dividing the two-dimensional space into four logical regions. In the logical quadrants defined in this embodiment, the region with a horizontal axis value greater than the dividing line and a vertical axis value less than the dividing line is defined as the first quadrant (high confidence, low noise region); the region with a horizontal axis value less than the dividing line and a vertical axis value less than the dividing line is defined as the second quadrant (low confidence, low noise region); and the region above the dividing line with a vertical axis value greater than the dividing line is uniformly defined as the third and fourth quadrants (physical defect region).

[0065] Specifically, the two-dimensional evaluation space consists of confidence level (X-axis) and intrinsic variance (Y-axis): the X-axis coordinate is the maximum value of the predicted confidence vector, and the Y-axis coordinate is the intrinsic variance scalar of the sub-quality branch output; the vertical axis threshold T_var: intrinsic variance judgment threshold, the default value is 0.2, and Y<0.2 is defined as "low variance region", Y≥0.2 is defined as "high variance region"; the horizontal axis threshold T_conf: predicted confidence judgment lower limit, the default value is 0.6, and X≥0.6 is defined as "high confidence region" and X<0.6 is defined as "low confidence region".

[0066] The orthogonal threshold diversion controller is configured within the edge computing node of the highway intelligent checkpoint system. Its internally preset two-dimensional evaluation space thresholds are: a prediction confidence threshold of 0.6 and an intrinsic variance threshold of 0.2. When a standard white sedan is captured in clear weather, the data point coordinates are calculated to be (0.96, 0.03), accurately falling into the first quadrant. The controller identifies it as a high-confidence sample and immediately executes a local retention instruction, writing the structured features of the image frame into the embedded database on the edge side. This is retained only as a local log and does not trigger network transmission.

[0067] Conversely, when the camera captured a modified farm vehicle carrying irregularly shaped cargo, the prediction confidence was only 0.35 due to a lack of such samples in the model library. However, the image itself was well-lit and blur-free, with an intrinsic variance scalar of 0.08, and the data point fell into the second quadrant. The controller determined it to be a "cognitive uncertainty sample" and immediately activated the 4G / 5G communication module to establish an AES encrypted channel to transmit this "clear but difficult-to-understand" high-value example back to the cloud. According to field test statistics, this strategy reduced the amount of data uploaded to the cloud daily from 500,000 captured images to approximately 25,000, but these 25,000 images contributed more than 90% to the model's accuracy.

[0068] Furthermore, for severely overexposed images caused by direct high beams from oncoming vehicles at night, or completely black frames resulting from sensor malfunctions, the intrinsic variance calculated by the sub-quality branch is as high as 0.95, with data points falling into the third or fourth quadrant. The controller determines this as an irreversible physical defect and generates a physical interception command within 10 milliseconds, directly clearing the cache stack of the frame data in DDR memory to prevent invalid scrap from occupying valuable end-side storage resources.

[0069] This invention clarifies the specific routing instructions for the first to fourth quadrants, establishing an unattended automated data governance closed loop that significantly reduces operating costs. By directly intercepting physically discarded fragments in the third / fourth quadrant at the edge, it prevents junk data from leaving the edge, saving several times the cloud storage space and expensive uplink bandwidth resources. Simultaneously, it accurately filters backflow samples with cognitive uncertainty in the second quadrant, allowing the cloud-based manual annotation team to focus on processing only the most valuable data, avoiding invalid annotation of simple samples or discarded fragments, and significantly improving the return on investment for manual annotation.

[0070] Furthermore, the cloud server receives data identified as high-value difficult examples uploaded from multiple edge devices and constructs an incremental training set by combining it with real labels generated by a manual annotation interface. The constructed incremental training set includes all newly annotated difficult example samples and their real labels within the current period. When the accumulated number of valid annotated samples reaches a threshold M (e.g., 10,000) and the time interval T (e.g., 7 days) since the last model update has exceeded, an incremental training process is triggered. The semantic-physical feature extractor is retrained using the incremental training set. While maintaining the uncertainty weighting constraint mechanism, the network parameters of the main semantic branch and the secondary quality branch are updated through the backpropagation algorithm. The updated model parameters are quantized and compressed, and distributed to the edge devices via wireless over-the-air download technology, replacing the pre-installed semantic-physical feature extractor in the edge devices.

[0071] Specifically, taking the city-wide intelligent traffic management cloud platform as an example, this platform connects 5,000 edge computing nodes distributed across various intersections. During the one-week operation period, the cloud server aggregated approximately 100,000 "high-value, difficult-to-identify" images uploaded by various nodes. These images mainly focus on the recently emerging category of "unmanned delivery vehicles," which were classified as quadrant data because the old model had never seen them before. After the cloud platform corrected the labels of these data to "special vehicles" through a manual annotation interface, it initiated the incremental training process.

[0072] While maintaining an active uncertainty-weighted constraint mechanism to continuously suppress background noise interference, after 50 epochs of backpropagation iterations, the new model's accuracy in recognizing unmanned delivery vehicles improved from the initial 35% to 92%. Subsequently, the server performed INT8 quantization compression on the updated model, compressing the original 180MB FP32 precision model to a 45MB INT8 precision model, while keeping the accuracy loss within 0.5%. Finally, the cloud platform utilized the network off-peak window from 3:00 AM to 4:00 AM to concurrently push the lightweight model package to all edge nodes via OTA technology. After the replacement was completed, the edge devices immediately possessed the ability to accurately identify unmanned delivery vehicles locally, eliminating the need to transmit them back as unknown and difficult examples, thus completing a closed-loop evolution of edge-cloud collaboration.

[0073] This invention describes the process of constructing an incremental training set and performing quantization updates in the cloud using challenging examples from the backflow side, thus building a cloud-edge collaborative data flywheel evolution system. This is not merely a static filter, but an organism with self-evolving capabilities. By continuously absorbing challenging examples mined from the edge, the model's recognition blind spots are continuously corrected. Combined with model quantization compression and OTA distribution technologies, front-end devices can quickly acquire more powerful recognition capabilities, thereby further improving the accuracy of subsequent data filtering and forming a virtuous cycle.

[0074] This invention constructs a feature extractor comprising a main semantic branch and a secondary quality branch, and introduces an uncertainty weighted constraint mechanism and an orthogonal threshold diversion controller. This decouples cognitive uncertainty from physical uncertainty, avoiding the inability to distinguish between rare samples and blurry images when filtering data solely based on predicted probabilities. This prevents the accidental deletion of high-value difficult examples or the mistraining of low-quality noise. The proposed solution quantifies physical noise through the secondary branch, orthogonally separating data in a two-dimensional space. This accurately removes invalid physical noise from massive datasets while retaining clear difficult examples, significantly improving the effective information density of the dataset used for downstream model training. This fundamentally solves the problem of garbage input and garbage output.

[0075] Example 2

[0076] This invention provides a data processing system for AI model data. Based on an edge-cloud collaborative architecture, the system aims to address the technical problem of being unable to effectively distinguish between high-value cognitive examples and low-quality physical noise when faced with massive data access, resulting in wasted computing power and low model iteration efficiency.

[0077] As one embodiment of the present invention, refer to Figure 1 A flowchart of a data processing method for AI model data, see reference. Figure 2 A schematic diagram of the structure of a data processing system for AI model data, referring to... Figure 3 A schematic diagram of an orthogonal threshold splitting strategy.

[0078] This system mainly consists of an edge data processing subsystem and a cloud model iteration subsystem, which are connected via an encrypted communication network. The edge data processing subsystem is typically integrated into an embedded SoC in a smart camera, vehicle computing unit, or industrial quality inspection robot; the cloud model iteration subsystem is deployed in a high-performance server cluster or data center.

[0079] The edge-side data processing subsystem includes a feature extraction module, a parameter optimization module, a traffic splitting control module, and a resource monitoring agent unit; the cloud-based model iteration subsystem includes a parameter optimization module, a data annotation interface, and a model distribution unit. The detailed structure and functions of each module are as follows: The feature extraction module is the core of the system's perception mechanism. It is configured to transform raw acquired data (such as video stream frames and image data) into high-dimensional vectors and quality assessment scalars that can be understood by computers. This module is implemented in hardware through an asymmetric dual-stream heterogeneous network topology, specifically comprising physically separated or logically isolated main semantic branch processing units and secondary quality branch processing units.

[0080] The main semantic branch processing unit employs a deep residual convolutional neural network architecture (such as ResNet-50). Its input is connected to the data acquisition interface, extracting high-dimensional semantic representations of the input data through multi-level nonlinear transformations (convolution, activation, pooling). The output of this unit is connected to a normalized exponential function (Softmax) layer, used to generate a prediction confidence vector representing the probability distribution of the data's category. During system operation, this unit is primarily responsible for understanding the content of the screen, such as identifying vehicle models or pedestrian types. Its computational load is significant, typically consuming the main computing resources of the NPU.

[0081] The sub-quality branch processing unit employs a lightweight shallow convolutional network architecture. Its network layer count is strictly designed to be less than half that of the main semantic branch (e.g., only 5-8 layers), and its parameter count is less than 5% of that of the main branch. This unit focuses on extracting low-level texture noise features from the data (such as motion blur, Gaussian noise, and overexposure). Its branch ends are connected to a global average pooling layer and a non-negative activation function layer (such as ReLU) to map the extracted features to a unique non-negative value—the intrinsic variance scalar.

[0082] The frequency-spatial hybrid feature extraction unit, as the core component of the sub-quality branch processing unit, integrates a Region of Interest (ROI) extractor, a block-based Discrete Cosine Transform (DCT) processor, and a feature fusion regressor. The ROI extractor receives the class activation map output from the shallow layer of the main semantic branch, generates a spatial attention mask through upsampling and binarization, thereby accurately locating the main target and shielding background interference. The DCT processor divides the main target region image selected by the mask into blocks and performs discrete cosine transform using a hardware-accelerated instruction set. This processor has built-in comparison logic circuitry to calculate the ratio of the transform coefficient energy to the DC component energy within a preset coordinate region of the transform matrix, and filters it in conjunction with a preset high-frequency noise energy upper limit threshold. Simultaneously, a standard quantization matrix is ​​introduced as a physical noise reference prior, and the number of zero-trace coefficients after quantization is counted to generate a frequency domain fingerprint vector. This process quantizes the physical sharpness of the image at the signal processing level.

[0083] The feature fusion regressor concatenates and fuses the weighted texture features in the spatial domain with the fingerprint vector in the frequency domain along the channel dimension, and generates the final intrinsic variance scalar through a fully connected layer. This scalar has undergone semantic focusing and frequency band energy calibration, which can resist misjudgment caused by background blurring.

[0084] The parameter optimization module in this system is divided into two parts: training state and inference state.

[0085] In cloud training mode, this module is the uncertainty weighted constraint mechanism, an algorithmic logic unit running on the training server, responsible for constraining and updating the parameters of the semantic-physical feature extractor during model iteration. Internally, the module includes a composite loss function calculation unit, containing task prediction error terms and regularization constraint terms. The dynamic weight adjustment subunit uses the reciprocal of the intrinsic variance scalar output of the secondary quality branch as the dynamic weight coefficient, performing division decay on the main task loss term. Its logic is: when the intrinsic variance scalar increases, it automatically reduces the weight of the prediction error generated by that sample on model parameter updates, thereby preventing the model from forcibly fitting noisy data. The damping constraint subunit simultaneously introduces a logarithmic term containing the intrinsic variance scalar into the loss function; when the model attempts to avoid task prediction errors by unrestrictedly increasing the intrinsic variance scalar, the value of this logarithmic term rises sharply and generates a strong back gradient, forcing the model to output high variance only when the data actually has physical defects. The orthogonal projection constraint subunit also calculates the cosine similarity between the main semantic feature vector and the secondary quality feature vector, forcing them to remain orthogonal in high-dimensional space, ensuring complete decoupling of semantic information and physical quality information.

[0086] The shunting control module is the decision-making center of the edge system, namely the orthogonal threshold shunting controller. This module receives the predicted confidence vector and intrinsic variance scalar output by the feature extraction module, and constructs a two-dimensional evaluation space in memory (the horizontal axis is confidence, and the vertical axis is variance). It then performs data shunting operations according to orthogonal discrimination rules. The module is further divided into the following sub-units: the resource monitoring agent unit polls and collects the hardware status parameters of the edge device at a millisecond frequency (e.g., every 50ms), specifically including the remaining battery percentage, NPU computing unit utilization, and the current uplink network. Network bandwidth (4G / 5G / Wi-Fi signal strength); The threshold dynamic drift unit internally stores a preset threshold drift curve, defining the negative correlation mapping relationship between the inherent variance judgment threshold and real-time computing load parameters; When the resource monitoring agent unit detects that the battery level is lower than the preset warning line (e.g., 20%) or the NPU utilization rate is too high (e.g., 90%), the unit will automatically adjust the inherent variance judgment threshold on the vertical axis along the curve (e.g., from 0.8 to 0.4), tightening the screening criteria and prioritizing the interception of more data to save energy; When uplink network congestion is detected, the unit... The unit improves the lower limit of prediction confidence for data in the second quadrant, reduces data backhaul, and prevents communication congestion. Furthermore, it features environmental baseline calibration, automatically adjusting thresholds to adapt to environmental noise drift caused by changes in day and night illumination by statistically analyzing data distribution within a sliding time window. The instruction generation and execution unit reads the quadrant placement of data points in the two-dimensional evaluation space in real time and generates corresponding control instructions: The first quadrant generates a local retention instruction, marking the data as a high-confidence sample and writing only the structured results to the local database; the second quadrant generates a cloud backhaul instruction, determining the data as a cognitive uncertainty sample and establishing an encrypted channel via the communication module to upload the data to the cloud's manually labeled interface queue; the third / fourth quadrant generates a physical interception instruction, determining that the data has irreversible physical quality defects and directly releasing the resources occupied by the data in the edge memory; before executing the cloud backhaul instruction, the desampling unit calculates the perceptual hash fingerprint of the current data and calculates the Hamming distance with historical fingerprints in the cache queue; if the distance is less than a threshold, it is considered a redundant sample and the upload is intercepted, only updating the statistical count, further saving bandwidth.

[0087] The cloud-based model iteration subsystem is responsible for achieving closed-loop model evolution using high-value data fed back from edge devices. The data receiving and incremental construction unit receives second-quadrant data uploaded from tens of thousands of edge devices, obtains real labels through a manual annotation interface, and constructs an incremental training set. The retraining execution unit uses the incremental training set to fine-tune the semantic-physical feature extractor. During this process, the system maintains an uncertainty-weighted constraint mechanism and uses the backpropagation algorithm to update network parameters, enabling the model to learn to recognize the original difficult examples. After training, the model quantization and distribution unit quantizes and compresses the model parameters to adapt to the inference accelerator of the edge chip; finally, it distributes the updated model to all edge devices via over-the-air (OTA) download technology, replacing the old version of the feature extractor and completing the system upgrade.

[0088] The system described in this embodiment achieves parallel processing of semantic perception and physical quality assessment under low computing power conditions at the edge through an asymmetric dual-stream heterogeneous network architecture, significantly reducing computational power consumption. By utilizing an orthogonal threshold diversion controller, the system can accurately decouple cognitive difficulties and physical noise in the two-dimensional assessment space, enabling high-value samples to be directed back while preventing junk data from leaving the edge. Combined with the dynamic threshold drift technology of the resource monitoring agent, the system has environmental adaptability and can automatically adjust its strategy to ensure business continuity when the network is congested or the power is low, greatly saving transmission bandwidth and cloud storage costs, and constructing an efficient edge-cloud collaborative data evolution closed loop.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data processing method for AI model data, characterized in that, include: Input the raw data into the preset semantic-physical feature extractor; The semantic-physical feature extractor includes a main semantic branch and a sub-quality branch. The main semantic branch is used to extract high-dimensional semantic features of the input data and generate a prediction confidence vector. The sub-quality branch is used to extract low-level texture noise features and generate an intrinsic variance scalar. The parameters of the semantic-physical feature extractor are constrained and updated through an uncertainty weighted constraint mechanism. The uncertainty weighted constraint mechanism uses the inherent variance scalar to perform division decay on the loss term of the main task, and simultaneously uses the regularization term to generate damping constraints on the parameter growth of the sub-quality branch, and iteratively optimizes the inherent variance scalar. The predicted confidence vector and the optimized intrinsic variance scalar are used as inputs to construct a two-dimensional evaluation space, which is then mapped to an orthogonal threshold shunting controller. The orthogonal threshold diversion controller includes orthogonal discrimination rules for dividing data quadrants, wherein the judgment threshold corresponding to the vertical axis is controlled by the real-time computing power load parameters of the end-side device; and executes the corresponding diversion command according to the quadrant of the data point in the two-dimensional evaluation space.

2. The data processing method for AI model data according to claim 1, characterized in that, The semantic-physical feature extractor employs an asymmetric two-stream heterogeneous network topology: The main semantic branch adopts a deep residual convolutional neural network architecture, which extracts high-dimensional semantic representations of the input data through multi-level nonlinear transformations. The network output is connected to a normalized exponential function layer to output the predicted confidence vector. The sub-quality branch adopts a lightweight shallow convolutional network architecture with fewer than half the number of layers as the main semantic branch. The branch ends are connected to a global average pooling layer and a non-negative activation function layer, which maps the extracted features to a non-negative intrinsic variance scalar. The main semantic branch and the secondary quality branch share shallow convolutional features at the input end, their parameters are independent in subsequent layers, and they run in parallel during the inference stage.

3. The data processing method for AI model data according to claim 2, characterized in that, The secondary quality branch employs a region-of-interest-guided frequency-spatial hybrid feature extraction mechanism. The process of generating the intrinsic variance scalar specifically includes: acquiring the class activation map output from the shallow network of the main semantic branch, performing upsampling and binarization to generate a spatial attention mask representing the location of the main target; using the spatial attention mask to extract the main region image from the input data and dividing the main region image into several pixel blocks; performing a block-based discrete cosine transform on each pixel block, calculating the ratio of the transform coefficient energy to the DC component energy within the preset coordinate region of the transform matrix, and filtering based on a preset high-frequency noise energy upper limit threshold; simultaneously introducing a standard quantization matrix as the physical... Using a noise reference prior, the number of transformation coefficients within the preset coordinate region where the quantized values ​​approach zero is statistically analyzed to generate a frequency domain fingerprint vector representing the physical clarity of the main target area. A convolutional layer is used to extract the original spatial texture feature map of the input data, and the spatial attention mask is multiplied element-wise with the spatial texture feature map to filter background region features and extract weighted texture features containing only the main target. The feature vector obtained by global average pooling of the weighted texture features is concatenated and fused with the frequency domain fingerprint vector along the channel dimension. The fused features are then input into a fully connected layer to regress and generate the intrinsic variance scalar, which has undergone semantic focusing and frequency band energy calibration.

4. The data processing method for AI model data according to claim 1, characterized in that, The uncertainty-weighted constraint mechanism's parameter constraint update process during the training phase includes: A composite loss function is constructed, including a task prediction error term and a regularization constraint term. The reciprocal of the sum of the intrinsic variance scalar output of the secondary quality branch and a preset smoothing minimum value is used as a dynamic weighting coefficient to weight and attenuate the task prediction error term. This dynamic weighting coefficient is negatively correlated with the intrinsic variance scalar. Simultaneously, a logarithmic term containing the intrinsic variance scalar is introduced into the loss function as a regularization constraint term. The magnitude of this regularization constraint term is monotonically increasing and positively correlated with the intrinsic variance scalar. By minimizing the composite loss function, the weight parameters of the main semantic branch and the secondary quality branch are jointly updated.

5. A data processing method for AI model data according to claim 1, characterized in that, The edge device integrates a resource monitoring agent that collects real-time computing load parameters of the edge device at millisecond frequency, including the remaining battery percentage, NPU computing unit utilization, and current uplink network bandwidth. The orthogonal threshold shunting controller stores a threshold drift curve, defining a negative correlation mapping relationship between the inherent variance judgment threshold and the real-time computing load parameters. When either the remaining battery percentage is lower than a preset warning line or the NPU computing unit utilization is higher than a preset saturation, the orthogonal threshold shunting controller lowers the inherent variance judgment threshold along the threshold drift curve. When the current uplink network bandwidth is detected to be congested, the orthogonal threshold traffic splitter controller raises the lower limit of the prediction confidence level for the second quadrant data.

6. The data processing method for AI model data according to claim 1, characterized in that, The orthogonal threshold diversion controller reads the coordinates of data points in the two-dimensional evaluation space in real time and generates control commands according to preset quadrant division rules: when a data point falls into the first quadrant consisting of a high prediction confidence interval and a low intrinsic variance interval, a local retention command is generated to mark the data as a high-confidence sample and store it in the local inference database of the edge device; when a data point falls into the second quadrant consisting of a low prediction confidence interval and a low intrinsic variance interval, a cloud backflow command is generated to mark the data as a cognitive uncertainty sample and establish an encrypted transmission channel with the cloud server through the communication module of the edge device, uploading the data as a high-value difficult example to the cloud's manual annotation interface queue; when a data point falls into the third and fourth quadrants consisting of a high intrinsic variance interval, a physical interception command is generated to determine that the data has irreversible physical quality defects and directly release the resources occupied by the data in the edge memory.

7. A data processing method for AI model data according to claim 1, characterized in that, The cloud server receives data identified as high-value difficult examples uploaded from multiple edge devices and constructs an incremental training set by combining it with real labels generated by a manual annotation interface. The semantic-physical feature extractor is then retrained using the incremental training set. While maintaining the uncertainty weighting constraint mechanism, the network parameters of the main semantic branch and the secondary quality branch are updated using the backpropagation algorithm. The updated model parameters are then quantized and compressed and distributed to the edge devices via wireless over-the-air download technology, replacing the pre-installed semantic-physical feature extractor in the edge devices.

8. A data processing system for AI model data, characterized in that, include: Feature extraction module: Inputs raw data into a preset semantic-physical feature extractor; The semantic-physical feature extractor includes a main semantic branch and a sub-quality branch. The main semantic branch is used to extract high-dimensional semantic features of the input data and generate a prediction confidence vector. The sub-quality branch is used to extract low-level texture noise features and generate an intrinsic variance scalar. Parameter optimization module: The parameters of the semantic-physical feature extractor are updated by means of an uncertainty weighted constraint mechanism. The uncertainty weighted constraint mechanism uses the inherent variance scalar to perform division decay on the loss term of the main task, and simultaneously uses the regularization term to generate damping constraints on the parameter growth of the sub-quality branch, and iteratively optimizes the inherent variance scalar. The shunting control module constructs a two-dimensional evaluation space by taking the predicted confidence vector and the optimized intrinsic variance scalar as inputs, and maps it to the orthogonal threshold shunting controller. The orthogonal threshold diversion controller includes orthogonal discrimination rules for dividing data quadrants, wherein the judgment threshold corresponding to the vertical axis is controlled by the real-time computing power load parameters of the end-side device; and executes the corresponding diversion command according to the quadrant of the data point in the two-dimensional evaluation space.

Images