34 results about "Dynamic modulation" patented technology

This invention discloses a river flow monitoring system and method based on ultra-high frequency side-scan radar. The data acquisition module is deployed on the riverbank to acquire multi-channel raw echo data required for river hydrological parameter inversion. The river flow monitoring cloud platform processes, analyzes, displays, and stores the data acquired by the data acquisition module, and sends corresponding instructions to the data acquisition module. This solution uses a database to correlate and store hydrological data, meteorological data, cross-sectional geometric data, and comparative measurement data in a time series manner, providing a solid data foundation for in-depth data mining, model optimization, and trend prediction. Furthermore, the data analysis considers the impact of complex tidal dynamic modulation in tidal river sections on river flow calculation. A tidal phase correction term is introduced into the index velocity method model, effectively eliminating the systematic bias caused by tidal dynamic modulation in the calculation.

This invention proposes a collaborative reconstruction method for multimodal images from unmanned aerial vehicles (UAVs) for all-weather perception. First, it uses an Adaptive Degradation Perceptual Block (ADPB) to perform spectral analysis on degraded visible light features to perceive the degradation type, and utilizes a learnable frequency mask to decouple features, providing a clear, high-fidelity structural prior for subsequent modal interactions. Next, it leverages the multi-scale decomposition capability of wavelet transform to decouple multi-frequency sub-band signals, and learns local contraction mappings within each sub-band through deep convolution, effectively preserving structural integrity while suppressing blur and noise. Finally, addressing the challenge of simultaneous degradation and cross-modal information fusion in multimodal images, it generates dynamic modulation factors in spatial and channel dimensions to guide adaptive complementary interaction and deep fusion of cross-modal features, strengthening the representation capabilities of global and local features. This significantly improves the reconstruction quality and multi-task generalization ability of the unified multi-task recovery model in complex scenarios.

The application discloses a video encryption method and device and a computer storage medium, and relates to the field of video encryption. The method comprises the following steps: calculating a spatial gradient amplitude and a motion vector field based on three continuous frames of an original video, calculating a motion activity to generate an activity tensor based on the two, constructing a complex seed-driven four-dimensional hyperchaotic mapping based on the activity tensor and a video resolution to generate three synchronous pseudo-random sequences, forming a heterogeneous encryption unit set based on the activity and a local Shannon entropy and an edge density product and combining dynamic modulation of the first pseudo-random sequence, mapping each heterogeneous unit into a virtual particle, constructing a chaotic potential field by the second pseudo-random sequence to simulate a virtual particle trajectory, and realizing pixel order rearrangement according to the trajectory, and generating a diffusion mask based on the third pseudo-random sequence to perform cross-frame XOR feedback diffusion on the result of the permutation. The application improves the security of video encryption.

The application discloses a gas recognition visual-olfactory fusion optoelectronic synapse device and a preparation method thereof, and belongs to the technical field of semiconductor devices. The optoelectronic synapse device comprises a substrate, a two-dimensional material floating gate heterojunction and source and drain electrodes. The heterojunction at least comprises a channel layer and a floating gate layer stacked thereon, and forms a Type-II energy band structure, and the interface has a defect state for carrier capture and release. The channel layer generates photo-generated carriers in response to light pulses, and the floating gate layer responds to gas stimulation and changes the charge state through gas adsorption / desorption; by applying a programmable light pulse stimulus to the channel layer, the adsorption or desorption process of gas molecules on the floating gate layer and the charge state thereof are dynamically modulated, a dynamic electrical response signal fusing visual and olfactory information is output, and gas recognition is realized. The application significantly improves the gas recognition accuracy from 52.24% of a single mode to 98.27% by using a dual-mode fusion feature, and realizes the synergy of miniaturization, low power consumption and high-precision perception.

The application discloses an intelligent bionic estrus induction and estrus detection robot system and a working method thereof, and the system comprises: an autonomous mobile inspection platform serving as a physical carrier of the estrus induction and estrus detection robot; an anti-habituation bionic excitation subsystem for generating a modulation control sequence and dynamically modulating a pre-recorded pig audio sample to output an anti-habituation bionic estrus induction audio, and combining a smell pulse to trigger a bionic estrus induction mechanism; a multi-modal perception and feature extraction subsystem for perceiving physiological data and behavioral reaction data of sows after being induced by the bionic estrus induction mechanism, and converting the data into multi-dimensional response features; and a large model diagnosis subsystem based on cross attention, for cross attention fusion of the multi-dimensional response features, obtaining an estrus diagnosis result according to the fused features, and driving the system to perform a closed-loop execution stage. The application effectively replaces a test estrus boar, and significantly reduces non-production days of breeding.

The application discloses an Au-WS2-Au van der Waals Schottky junction type photoelectric sensing unit, preparation and application, and belongs to the technical field of semiconductor devices. The sensing unit is composed of a substrate, a two-dimensional material layer transferred by a dry method and an Au electrode, the two-dimensional material is preferably WS2, the Au electrode forms a van der Waals contact with the two-dimensional material and constitutes a Schottky junction; the structure can effectively inhibit Fermi level pinning effect, realize dynamic modulation of a Schottky barrier, and produce characteristic reverse-bipolar light response and stable spectral modulation varying with wavelength under the regulation of a gate; the occupied area is only about 73 mu m 2 , and the sensing unit can accurately reconstruct monochromatic and broadband spectra in the 380-680 nm visible light band, with an average reconstruction accuracy of 0.746 nm and a spectral resolution better than 10 nm. The sensing unit can be applied to a miniature computing spectrometer, effectively solves the technical defects of complex preparation and difficult regulation of existing devices, and meets the requirements of miniaturization and integration detection.

This invention provides a method and system for audio-visual instance segmentation based on audio-visual fusion, comprising: S1: acquiring a video sequence to be processed and its corresponding audio signal, and extracting visual features of the video sequence and audio features of the audio signal; S2: spatially modulating the visual features based on the audio features to enhance the response intensity of the sound-producing region in the visual features, thereby obtaining enhanced visual features; S3: semantically aligning and reverse-guiding the audio features based on the enhanced visual features, thereby obtaining semantically enhanced object-encoded features; S4: aggregating the enhanced visual features and object-encoded features, and decoding to generate segmentation and tracking results for the sound-producing target in the video sequence; S5: outputting the segmentation and tracking results. This invention constructs a pixel-level and object-level dual-layer audio-visual fusion mechanism to achieve dynamic modulation of visual representation by audio cues and reverse guidance of audio localization by visual semantics, thereby accurately locating and continuously tracking sound-producing instances in highly dynamic and complex scenes.

This invention discloses a diesel yield prediction method and system based on a time-delay-aware Transformer model, belonging to the field of industrial big data processing technology. It determines the effective time-delay window for each dynamic process variable through time-delay analysis constrained by physical mechanisms. Then, it embeds context vectors into static process variables and encodes dynamic process variables to generate dynamic feature sequences. A dynamic modulation gate generated from the context vectors is used to adaptively weight and modulate the dynamic feature sequences. Simultaneously, a global causal mask matrix is ​​constructed based on the effective time-delay windows of each variable and embedded with Transformer attention calculations, forcing the current query to access only historical moments within the effective time-delay window. The output is a temporal implicit representation conforming to the physical time-delay causality law, thereby predicting diesel yield. This method embeds the process mechanism knowledge of the refining process into a structured prior form within a deep learning architecture, achieving high-precision, high-robustness, and high-interpretability online rolling prediction of diesel yield.

This invention relates to the field of photoelectric signal energy conditioning technology and discloses an image enhancement method for wound assessment, comprising: acquiring a multi-channel photoelectric raw signal stream under an unsteady excitation source; constructing a reference energy response vector in the spectrum and power coupling vector space; determining the energy overload characteristic value based on the vector displacement deviation of the signal and generating an energy distribution coefficient through reverse negative feedback logic; and dynamically modulating the power amplitude component according to the coefficient. Under the constraint of locking the characteristic energy of the spectrum component, this invention achieves precise suppression of transient amplitude surges and solves the technical problems of signal-to-noise ratio degradation and source-end bandwidth impedance characteristic truncation under unsteady conditions.

The application discloses a marine environment multi-parameter coupling equipment fault diagnosis method, and belongs to the technical field of marine equipment state monitoring and fault diagnosis. Real-time observation values of multiple environment parameters are collected, a multi-parameter coupling strength index is calculated, a forbidden edge set and a necessary edge set are constructed, a hard constraint is imposed on data-driven causal diagram learning, a directed acyclic graph is output, a hierarchical dynamic Bayesian network including an environment layer, a degradation layer and a fault layer is constructed, and the multi-parameter coupling strength index is introduced into the state transition probability of the degradation layer to realize dynamic modulation; a multi-fault concurrent scene is decoupled and identified, independent concurrent faults and composite faults are distinguished, and component confidence is output, and online continuous learning is realized through a Bayesian conjugate increment method. The application effectively solves key problems such as multi-environment parameter coupling effect quantization, fusion of physical constraints and data driving, and sensor signal quality self-adaptation, and has long-term self-adaptation capability to service environment changes and new fault types.

The application discloses an endoscope depth estimation method and system based on dynamic modulation Mamba, and belongs to the technical field of medical image processing.The application combines Mamba and dynamic vector driven parameter regulation.The Mamba is responsible for efficiently capturing the spatial sequence correlation of the endoscope image, and solves the problems that the global context is difficult to model in the traditional model and the calculation complexity of the Transformer is high;the dynamic vector and the learnable projection layer enable the state update of the Mamba to adapt to the feature difference of different endoscope images, and improve the robustness of the model to the low-texture and uneven-light scene.Meanwhile, the design of the ViT backbone network ensures the effectiveness of feature extraction and the accuracy of depth regression, and forms a complete end-to-end depth estimation system.

This invention relates to the field of computer vision and image processing technology, specifically disclosing a target tracking method and system based on global semantics and spatial inductive bias. The method includes: performing image patch embedding on a template image using an image patch embedding module, and extracting features through a Transformer encoder to obtain a deep template label sequence; performing global average pooling on the deep template label sequence to obtain a global template semantic vector; performing image patch embedding on a search image using the image patch embedding module, and extracting features through a Transformer encoder to obtain a search feature sequence; and embedding a gated global template modulation module into the tracking model to perform channel-level dynamic modulation of the global template semantic vector and the search feature sequence with extremely low computational overhead, thereby achieving adaptive suppression of channel-level background noise and semantic purification of early features, effectively alleviating the problem of early semantic loss in lightweight asynchronous frameworks.

This invention relates to an adaptive anti-sunlight interference infrared detection method and system, belonging to the field of infrared detection technology. The system includes: an infrared transmitter, an infrared receiver, a signal preprocessing unit, a bandpass filter unit, a digital phase-locked loop unit, and a controller unit. The controller unit is electrically connected to the infrared transmitter, infrared receiver, signal preprocessing unit, bandpass filter unit, and digital phase-locked loop unit, and includes: a dynamic modulation module, a coherent processing module, an encoding and demodulation module, a feature extraction module, an intelligent decision-making module, and a detection output port. This invention, through a collaborative mechanism of dynamic modulation, coherent demodulation, and multi-feature verification, ensures that only signals emitted by the system itself can trigger detection. This allows the system to proactively adapt to environmental changes and accurately identify the true signals emitted by the system and reflected by obstructed objects. Thus, it achieves extremely high reliability and extremely low false alarm rate object detection in various harsh light environments, including direct sunlight.

The application belongs to the technical field of image fusion, and proposes an image fusion controllable modal modulation method based on non-classifier guidance: firstly, a double-branch neural network containing an unconditional reconstruction branch and a conditional fusion branch is constructed, the unconditional branch is used for high-quality fidelity reconstruction of the source image to suppress pre-background noise in the absence of conditional modal guidance, and the conditional branch is used for extracting features from the source image and the conditional modal image, and performing dynamic modulation and fusion in a multi-scale space; secondly, a three-stage joint training strategy is used to train the constructed double-branch neural network; then, in the inference stage, the source image, the conditional modal image and the set guide scale factor are received, unconditional prediction and conditional prediction are performed respectively, and continuous controllable modal injection without retraining is realized through a linear extrapolation formula, and the final fused image is output. The application effectively improves the precision and flexibility of the existing image fusion algorithm in the integration of heterogeneous modal features.

This invention discloses a medical image segmentation method based on KAN dual-branch and MIFA-HiLo collaborative optimization, belonging to the field of artificial intelligence and medical image processing technology. It aims to solve the technical problems of insufficient nonlinear expression capability, imbalance between local and global feature modeling, and insufficient multi-scale information fusion in existing methods. The method includes: acquiring a preprocessed image tensor; constructing a dual-branch encoder containing convolutional and KAN-Transformer branches; constructing a MIFA multi-scale integrated feature aggregation module to achieve dynamic modulation and fusion of local and global features through a bidirectional cross-gating mechanism; constructing a multi-scale progressive decoder to progressively restore spatial resolution; constructing a HiLo feature optimization module to perform high-frequency enhancement and low-frequency recalibration of features through channel partitioning; and finally outputting a pixel-level probability prediction map.

The application discloses a laser electric arc hybrid additive manufacturing system based on trajectory-power dynamic regulation and a working method thereof. The system comprises a laser unit, a beam shaping unit, a deformation optical unit, an electric arc unit, a first wire feeding unit, a second wire feeding unit, a substrate unit, a robot unit, a monitoring and control unit, a laser swing mode unit and a laser power modulation curve unit. The laser energy distribution and wavefront are dynamically modulated through the beam shaping and deformation optical unit, and the laser scanning trajectory and the laser power are cooperatively controlled to construct controllable energy distribution and temperature gradient within the dot matrix unit scale. Meanwhile, an electric arc heat source is introduced as an auxiliary heat input to form a complementary hybrid heat source with the laser to stabilize the molten pool shape and prolong the effective existence time. The system also realizes controllable deposition of single material, multi-material or gradient material dot matrix structure through independent control of the double wire feeding units. The corresponding working method realizes stable forming of the dot matrix structure through point-by-point and layer-by-layer trajectory-power cooperative regulation. The application can effectively improve the consistency, stability and metallurgical quality of the dot matrix structure forming, and is suitable for manufacturing of lightweight, high-performance and complex dot matrix structure parts.

The application relates to the technical field of radar target detection and data fusion, and discloses an airport super-low-altitude clutter suppression and target detection method based on millimeter wave radar blind filling, which comprises the following steps: dividing a three-dimensional voxel set, controlling a millimeter wave radar network to construct a voxelized dynamic Doppler profile baseline; performing space-time registration and compensation on asynchronous point data of a main radar network and the millimeter wave radar network; calling the voxelized dynamic Doppler profile baseline to output millimeter wave target measurement data, calculating environmental Doppler information entropy to dynamically modulate a measurement noise covariance matrix; calculating and outputting a heterogeneous joint track state and predicting a target track, and stopping the update of the voxelized dynamic Doppler profile baseline corresponding to an affected three-dimensional voxel subset. The technical scheme that the three-dimensional voxel environmental Doppler information entropy is calculated to dynamically modulate the measurement noise covariance matrix is adopted, so that the technical effect of adaptively adjusting the measurement weight of a fusion filter is achieved.

The application discloses a kind of remote sensing image feature optimization method and device based on physical base, belong to remote sensing image processing technical field.The method is first obtained the deep feature representation of remote sensing image;Subsequently, deep feature mapping is constructed as feature physical field representation;Again, a set of physical base functions for describing different spatial structure characteristics is constructed, and the feature physical field representation is mapped and projected to different physical base spaces, realizing feature decomposition based on physical base;Then, according to the spatial context information of feature, each physical base feature component is adaptively modulated, and the modulated feature component is dynamically reconstructed and fused to generate enhanced feature representation.The application realizes effective optimization of remote sensing image feature representation by introducing physical base constraint and dynamic modulation mechanism, can enhance the representation ability of multi-scale structure, complex boundary and detail information, and has good universality and scalability.

This invention discloses an image super-resolution method based on feature attribute dynamic modulation nonlinear quantization. The steps are as follows: S1, acquire the super-resolution model and the image set for training; S2, construct a bit-width dynamic modulation module, using the frequency domain spectral entropy calculated through the forward input tensor as a feature attribute index to achieve dynamic bit-width modulation; S3, construct an activation value quantization module, performing nonlinear compression transformation on the forward input tensor and outputting the quantized activation value; S4, construct a weight value quantization module, outputting the quantized weight tensor through uniform quantization of the channel-level weight tensor; S5, perform convolution operation between the quantized activation value and the quantized weight tensor; S6, train the super-resolution model improved by step S5 using the image set. This method achieves a peak signal-to-noise ratio / structural similarity superior to existing mainstream training quantization methods at low bit depths, visually eliminates artifacts, restores a clear structure, and does not increase additional inference computation overhead.

This invention relates to the field of functional materials technology, specifically to a force-induced chiral liquid crystal compound, its preparation method, and its applications. The invention involves esterification, reduction, bromination, cyano substitution, and Knaufenberg-Gale condensation reactions to link a chiral alkyl chain-modified cyanostylene atom to trimesoaldehyde, yielding a liquid crystal molecule with a disk-like structure. This liquid crystal compound forms a non-helical columnar liquid crystal structure when unstressed, but can transform into a thermodynamically more stable monochiral helical columnar liquid crystal structure upon application of mechanical force, exhibiting significant circular dichroism and circularly polarized emission characteristics, with an emission asymmetry factor as high as 0.3. The force-induced chiral liquid crystal compound provided by this invention allows for dynamic modulation of chiral optical signals through simple mechanical stimulation, offering simple operation, direct response, and good reversibility.

The application discloses a power grid broadband oscillation positioning method and system for sparse dynamic graph space-time deduction, and the method comprises the following steps: reconstructing an enhanced graph topology based on electrical impedance transition probability, performing mask zero shielding on the features of unobservable nodes to overcome the topological fault caused by local observation; constructing a dynamic mode deduction module fusing multiple dimensions of state, projecting an oscillation sequence to a hidden manifold space to perform physical dynamic deduction, and extracting high-order dynamic features reflecting the evolution law of the system; combining node static features, inputting the high-order dynamic features into a multi-layer neighborhood state synchronous space-time aggregation graph network, introducing a cross-dimensional space-time feature self-adaptive scale balance mechanism for dynamic modulation, deeply fusing space-time information to obtain node comprehensive features; and using a classifier to realize oscillation tracing by combining multi-source constraint joint optimization. The application breaks through the limitation of a conventional black box model, and significantly improves the precision and robustness of oscillation positioning under sparse data.

The application relates to a domain adaptive target detection method based on dynamic data driving and consistency constraint, and belongs to the technical field of target detection. The method comprises the following steps: a target detection model containing a teacher-student network is constructed, and labeled source domain images and unlabeled target domain images are input; a semantic prediction consistency loss composed of semantic context consistency loss and prediction distribution consistency loss is calculated; a dynamic modulation factor is calculated based on the loss, a frequency domain filter cutoff frequency threshold is adjusted, and frequency domain dynamic processing is performed on the source domain images; the processed source domain images are input into the student network to calculate a supervised detection loss, and a total loss function is constructed in combination with the semantic prediction consistency loss; and the model is iteratively optimized based on the total loss function, the student network parameters are updated through back propagation, and the teacher network parameters are updated through exponential moving average. While improving the robustness and accuracy of cross-domain detection, no additional inference cost is required, and the engineering application value is high.

This invention belongs to the field of nighttime target image processing technology and discloses a closed-loop nighttime target tracking method and apparatus based on state feedback. The method involves feeding a low-light image sequence into a low-light perception enhancement network, generating enhanced images based on Retinex decomposition and structure-preserving feature enhancement; inputting the enhanced images into a target tracking network, outputting the target position based on the template-search interaction of a Transformer, and generating a target attention map and tracking confidence; feeding the target attention map and confidence into a tracking state feedback network, encoding target feedback modulation information, and feeding it back to the target modulation fusion module. The dynamic modulation enhancement process strengthens the target region and suppresses background noise, achieving a closed-loop nighttime target tracking process of "enhancement—tracking—feedback—re-enhancement". This invention can improve tracking stability, noise resistance, and robustness in complex low-light scenarios, and is suitable for low-light vision tasks such as nighttime surveillance, UAV nighttime inspection, and autonomous driving replication perception.

This invention discloses a cross-individual EEG real-time early warning method and system for construction site operation scenarios. The method first collects and preprocesses EEG signals in the construction site environment, then performs cross-individual covariance alignment and feature whitening on the preprocessed EEG signals. Next, a cross-individual fatigue separability perception and dynamic modulation module is constructed to evaluate the separability of EEG features in each time window after feature whitening and adaptively modulate them. Then, a hybrid deep learning network is proposed to extract temporal features from the adaptively modulated EEG signals and predict fatigue states. Finally, the hybrid deep learning network is trained, followed by small-sample fine-tuning of cross-individual transfer learning. After fine-tuning, a smoothing mechanism is used to output the final real-time EEG early warning result. This invention can automatically learn effective EEG signal representations and maintain high classification accuracy even under high noise conditions.

The present application relates to a kind of impulse neural network small sample target detection systems, including impulse prototype construction module, prototype guided feature reweighting module, prototype matching detection head, Episode meta-training controller, wherein: the impulse prototype construction module is used to extract class prototype representation, carry out timing weighted aggregation, obtain the class prototype vector of fusion timing information;The prototype guided feature reweighting module is used to carry out dynamic modulation to query pulse feature;Carry out multi-prototype fusion, feature reweighting, so that query feature is close to the prototype of correct class;The prototype matching detection head includes boundary box regression branch and prototype matching classification branch;Prototype matching classification branch carries out classification decision, and outputs class probability;The Episode meta-training controller is used to carry out the sampling and training control of Episode task;For each Episode, maintain independent batch normalization statistics, optimize training process.The present application can realize the efficient target detection of impulse neural network in small sample scene.

The application relates to the field of remote sensing image processing and intelligent interpretation, and specifically discloses a remote sensing small target intelligent detection method and system based on Mambaout reparameterized convolution. The method first extracts multi-level features through a convolution backbone network embedded with a Mambaout-RepLK feature enhancement module, and realizes dynamic modulation of the features and spatial context modeling through gated convolution and large kernel convolution; secondly, the spatial and semantic relationship of multi-scale features is enhanced through an adaptive sparse self-attention feature interaction module; subsequently, global semantic information integration is realized through a multi-scale feature fusion module based on a Transformer coding structure; finally, the end-to-end prediction of target categories and bounding boxes is completed through an encoder-decoder type detection module. The application effectively reduces the calculation complexity while ensuring the detection accuracy, significantly improves the detection performance and model robustness of small targets in remote sensing images, and has a good engineering application prospect.

The application discloses a kind of original ternary computing systems and methods with environmental synergy capability, to build a live, the evolution of computing net, belong to novel computer architecture technical field.The application is ternary (+1,0,-1) as minimum logic unit, with integrated exponential divergence, logarithmic convergence, complex rotation and zero calibration four operations core integrated operation unit as unified operation engine, and with the specified fractal period of pi as the basis scheduling rhythm instead of traditional fixed clock.System at the end of each reference operation period, through optical negative feedback loop, output light field direct current bias fluctuation is controlled within ±0.1dB, realizes internal state reset, supports continuous 10^5 times or more without cumulative error operation cycle, to build the basic framework with inherent resilience and evolution ability.Its reserved environmental perception interface can be used as optional extension, by accessing external sensing module, dynamically modulate its internal rhythm, so that it has senior ability of coevolution with external environment.

The application discloses a three-dimensional Gaussian movable human avatar modeling method based on graph structure optimization, and belongs to the field of computer vision and three-dimensional reconstruction. The application firstly constructs a static dressing three-dimensional Gaussian human avatar template in a standard pose space, so as to provide stable geometric and material priors for subsequent dynamic modulation. Secondly, a multi-level pose graph is constructed based on the kinematic structure of the human body, feature propagation and aggregation are performed through a hierarchical graph neural network, and dynamic modulation of the appearance driven by the decoupled components is realized. Finally, a material perception skinning graph is constructed, geometric deviations and material characteristics are fused on the local neighborhood graph structure, and the skinning weight is structurally propagated and adaptively corrected through a gating mechanism, so that the pose transformation and rendering of the three-dimensional Gaussian element are completed. Through the introduction of the graph structure optimization mechanism, the appearance consistency and local deformation stability of the dressed human body under unseen poses are effectively improved, the pose generalization ability and visual realism of the movable human avatar are enhanced, and the application can be widely applied to the fields of virtual digital people, augmented reality, interactive rendering and the like.

The application relates to a point cloud denoising method and system based on residual enhancement and conditional modulation ScoreNet, which comprises the following steps: obtaining three-dimensional noisy point cloud data and preprocessing the same to obtain standardized point cloud data; inputting the standardized point cloud data into a point cloud feature extraction network, obtaining an initial feature map after feature initialization, and outputting point cloud fusion features through multi-stage feature aggregation and local feature enhancement; constructing a hidden conditional vector, inputting the hidden conditional vector and the point cloud fusion features into a conditional modulation residual ScoreNet architecture, dynamically modulating a predicted noise gradient score through a conditional modulation residual block; and denoising and updating each point of the noisy point cloud according to the noise gradient score to obtain a pure point cloud. The application alleviates the gradient vanishing problem of a deep network through multi-stage feature aggregation and local enhancement, dynamically and adaptively modulates gradient prediction by taking the point cloud fusion features as a hidden conditional vector, overcomes the defect of insufficient prediction accuracy of a fixed conditional vector, and exhibits stable denoising effects under different noise levels.

The application discloses an automatic driving perception method and system based on dynamic neural operators and physical evolution, and relates to the technical field of automatic driving. Specifically, the method comprises the following steps: extracting a multi-scale initial feature map of an environment image; mapping the initial feature map to a frequency domain to obtain a frequency domain feature, modulating the frequency domain feature by using a physical attenuation mask and a dynamic modulation parameter to simulate the evolution process of the feature in a continuous physical field, and restoring the feature to a spatial domain to obtain a local physical evolution feature; using a linear complexity attention mechanism to capture cross-region dependencies to generate global semantic features; based on the scene complexity difference, performing spatial-level weighted fusion on the local physical evolution feature and the global semantic feature to obtain fused perception features; and outputting a perception result by using the fused perception features. The application aims to introduce physical equation constraints to improve the perception robustness, and to realize global collaboration at a low computational cost, which is suitable for real-time environment perception of a vehicle-mounted edge computing platform.