CN-121981905-A - Multi-mode data enhancement method based on light weight

CN121981905ACN 121981905 ACN121981905 ACN 121981905ACN-121981905-A

Abstract

The invention relates to the technical field of electric digital data processing and discloses a light-weight-based multi-mode data enhancement method which comprises the steps of obtaining a multi-mode feature tensor, extracting vector orthogonal projection scalar and determining novelty, updating a global covariance matrix when the novelty is larger than a redundancy threshold value and keeping the global covariance matrix in a registered state when the novelty is not larger than the redundancy threshold value, extracting main diagonal variance components to determine a differential modulation coefficient, calculating a second-order rectangular manifold projection operator according to the global covariance matrix and calibrating the operator by utilizing the differential modulation coefficient, generating a structured disturbance vector for random noise orthogonal projection by utilizing the calibrated operator and superposing the structured disturbance vector to multi-mode feature tensor to output enhancement features.

Inventors

HUANG YUFENG
SU QIGUI
WU JINGXING
WANG PENG
XU ZHUANG
DU YONGHENG
XIAO JING
Guo huanyu
ZHANG ZUOQIANG
WEI GANGQIANG
LIU DAIDI
NIE RONG

Assignees

长沙谱蓝网络科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (9)

1. The multi-mode data enhancement method based on light weight is characterized by comprising the following steps of: step S1, acquiring a multi-mode characteristic tensor; S2, extracting vector orthogonal projection scalar quantities of the multi-mode feature tensors in a feature space, and determining novelty according to the vector orthogonal projection scalar quantities; step S3, when the novelty is larger than a preset redundancy threshold, executing update processing of the global covariance matrix to obtain an updated global covariance matrix; S4, extracting a variance component of a main diagonal position of a global covariance matrix in a current registering state, and determining a differential modulation coefficient of each characteristic channel according to a negative correlation relation between the variance component and the semantic strength of each characteristic channel, namely, the greater the self-variance component is, the weaker the semantic rigidity strength of the characteristic channel is, and the flexible semantics are presented; S5, calculating a second moment manifold projection operator by using the global covariance matrix, and calibrating the second moment manifold projection operator by using the differential modulation coefficient; s6, performing orthogonal projection on the random noise vector by using the calibrated second-order moment manifold projection operator to generate a structured disturbance vector aligned with the multi-modal feature tensor semantically; And S7, superposing the structured disturbance vector on the multi-modal feature tensor, and outputting the enhanced multi-modal feature.
2. The method for enhancing multi-modal data based on light weight according to claim 1 is characterized in that step S2 specifically comprises projecting a multi-modal feature tensor at a current moment to an orthogonal basis space formed by historical moment features to obtain a vector orthogonal projection scalar representing residual intensity, comparing the vector orthogonal projection scalar with a preset redundancy threshold, and if the vector orthogonal projection scalar is larger than the preset redundancy threshold, judging that novelty accords with a trigger criterion of update processing.
3. The method for enhancing multi-mode data based on light weight according to claim 1, further comprising a mask truncation step based on a feature accumulation contribution rate, wherein the method comprises a step S31 of executing feature value decomposition on a global covariance matrix in a current registering state to obtain a feature vector base set, a step S32 of arranging the feature vector base set according to the sequence of feature values from large to small to calculate an accumulation feature contribution rate, a step S33 of determining a truncation position when the accumulation feature contribution rate reaches a preset coverage threshold and rejecting non-core semantic noise base after the truncation position to construct a truncated projection matrix, and a step S34 of updating a second-order rectangular manifold projection operator by using the truncated projection matrix.
4. The method for enhancing multi-mode data based on light weight according to claim 1, wherein step S4 specifically comprises the steps of extracting self-variance components of corresponding characteristic channels in a global covariance matrix, establishing a positive correlation mapping relation between the self-variance components and differential modulation coefficients, distributing tiny differential modulation coefficients based on rigid semantic channels with the self-variance components in a range from 0.1 to 0.3, and distributing sufficient differential modulation coefficients based on flexible semantic channels with the self-variance components larger than a 0.7 threshold.
5. The method of claim 4, wherein the logic for calculating the differential modulation factor follows the formula: , wherein, For the differential modulation factor of the ith characteristic channel, Is the self-variance component of the ith element on the main diagonal of the global covariance matrix, and Located in the formula molecular terms to ensure Along with it Increasing with increasing, lambda being a preset intensity adjustment factor, Is a constant for maintaining numerical stability.
6. The method for enhancing multi-mode data based on light weight according to claim 1, wherein step S6 further comprises a momentum orientation projection step based on characteristic time sequence residual errors, wherein step S61 is implemented by repeatedly using characteristic difference vectors of adjacent time slices as disturbance driving sources, the time interval of the time slices is not more than 100ms, step S62 is implemented by carrying out weighted fusion on the disturbance driving sources and random noise vectors to generate driving operators, and step S63 is implemented by utilizing a second-order moment manifold projection operator to project the driving operators, so that the generated structured disturbance vectors are anchored to a data evolution track.
7. The method for enhancing multimodal data based on light weight as claimed in claim 6, further comprising a step of weight attenuation based on a Fu Luo Beini Usne 'S norm driving, wherein the step of calculating algebraic difference between a global covariance matrix and a local sliding window matrix at the current time is performed in the step of S71, the step of extracting the Fu Luo Beini Usne' S norm of the algebraic difference and representing the structural inversion degree of the multimodal feature distribution by using the Fu Luo Beini Usne 'S norm is performed in the step of S72, and the historical weight attenuation processing is performed when the Fu Luo Beini Usne' S norm exceeds a preset mutation threshold.
8. The method for enhancing multi-modal data based on light weight according to claim 1, further comprising a closed loop feedback step based on projection residuals, wherein the Euclidean distance between the multi-modal feature tensor and the projection vector of the multi-modal feature tensor on the second-order rectangular manifold projection operator is calculated, the projection residuals are obtained, the uncertainty of the data is represented by the projection residuals, and the proportionality coefficient of the structured disturbance vector when the multi-modal feature tensor is overlapped is dynamically adjusted according to the uncertainty of the data.
9. The method for enhancing multimodal data based on light weight as recited in claim 1, wherein the act of intercepting the update process in step S3 is specifically to monitor status bits of novelty through underlying conditional branch instructions, and directly skip write operation tasks for covariance accumulation registers when the status bits indicate that the novelty is not greater than a predetermined redundancy threshold, so that the global covariance matrix does not undergo state transition in a current processing cycle.

Description

Multi-mode data enhancement method based on light weight Technical Field The invention belongs to the technical field of electric digital data processing, and particularly relates to a multi-mode data enhancement method based on light weight. Background In the current process of processing electric digital data, the enhancement of perception precision by fusing a visual image with multi-modal feature vectors acquired by laser point cloud and the like belongs to an industry common means, enhancement processing is generally needed in a feature space to expand sample coverage in order to adapt a perception model to distribution characteristics under different physical environments, multi-modal features present specific statistical correlation in a joint space, the correlation determines a manifold structure of the feature tensor, and the manifold structure is a basis for ensuring that the enhancement features have semantic rationality, however, an edge side processing unit is limited by limited floating point operation capability and memory bus bandwidth, and when processing a feature stream input at a high frequency, balance between instantaneity and computational energy efficiency is difficult to achieve. In the prior art, common strategies aiming at feature enhancement comprise a reconstruction scheme based on neural network generation and a simplified scheme based on random disturbance, wherein the reconstruction scheme relates to complex nonlinear transformation and high-intensity calculation iteration, the generated calculation load is far away from a real-time processing threshold value of a chip at the super-edge end, the simplified scheme is extremely easy to cause feature distribution to deviate from an original manifold track due to neglecting the coupling constraint of statistical moment among modes, so that mode mismatch and semantic distortion are caused, even if deviation is attempted to be controlled by adjusting disturbance coefficients, the real-time perception of a bottom data distribution rule is often caused to be caused to show serious statistical hysteresis of a system in a dynamic environment, the control method cannot realize high-precision distribution modeling under limited resources, the perception performance is limited, for example, a data analysis method, a device, equipment and a medium fusing large numbers are disclosed in the Chinese patent application with a CN121365352A, the noise attenuation weighting is carried out by utilizing large number LLN convergence characteristics, the scheme is extremely dependent on the fact that the steady-state minimum sample quantity is used as an operation premise that the statistical sample quantity is perceived in a sub-severe-level real-time scene, the statistical weight is easily generated, the statistical quantity is different from the low-level statistical delay is caused by the adjustment of the transient state error coefficient, the transient state has no-scale coefficient, the current has a large-scale error and the calculation method is easily caused to be influenced by the fact that the mass-critical-scale has a large-of the calculation of the transient-scale-caused by the transient-caused error of the algorithm has a large-caused error, and the mass-caused by the mass-caused error-caused by the method. Therefore, how to construct a feature enhancement mechanism with statistical constraint and environment awareness capability under lower calculation cost becomes the technical problem to be solved by the invention. Disclosure of Invention The invention provides a multi-mode data enhancement method based on light weight, which comprises the following steps: step S1, acquiring a multi-mode characteristic tensor; S2, extracting vector orthogonal projection scalar quantities of the multi-mode feature tensors in a feature space, and determining novelty according to the vector orthogonal projection scalar quantities; step S3, when the novelty is larger than a preset redundancy threshold, executing update processing of the global covariance matrix to obtain an updated global covariance matrix; S4, extracting a variance component of a main diagonal position of a global covariance matrix in a current registering state, and determining a differential modulation coefficient of each characteristic channel according to a negative correlation relation between the variance component and the semantic strength of each characteristic channel, namely, the greater the self-variance component is, the weaker the semantic rigidity strength of the characteristic channel is, and the flexible semantics are presented; S5, calculating a second moment manifold projection operator by using the global covariance matrix, and calibrating the second moment manifold projection operator by using the differential modulation coefficient; s6, performing orthogonal projection on the random noise vector by using the calibrated second-order moment manifold projection operat