28results about How to "Improve recognition stability" patented technology

The invention discloses a traditional Chinese medicine residue sorting and quality grading method based on machine vision recognition. The method comprises the steps that an image sequence is collected, and exposure, white balance gain and light source driving quantity are recorded; a standardized image sequence is obtained through preprocessing, and a working condition drift tensor is constructed through batch difference; segmenting the target area, constructing an equivalent refraction-scattering weight map based on a particle size distribution parameter, a fiber orientation anisotropy parameter and a water-containing scattering indication parameter, and generating a physical token sequence; adopting improved VMama asymmetric scanning to obtain an anisotropic state characterization sequence; a prediction entropy is calculated, a physical consistency loss is constructed through a working condition drift tensor alignment residual error, and a self-adaptive normalized parameter set is obtained through TENT online updating; and reasoning to obtain a target sorting category and a target quality grade, and generating a shunting control instruction and grading record data for sorting and grading of a conveying line. According to the method, unmarked cross-working-condition sorting and consistent grading are achieved.

The application discloses a cement-based cementitious material phase identification method, which obtains a BSE image, a qualitative element surface spectrum and a quantitative element surface spectrum of a cement-based cementitious material sample, carries out pore extraction, superpixel division, PHATE dimension reduction, GMM clustering and multi-view interactive platform correction, derives an EDS data correction table, constructs an element fingerprint library, carries out initial probability prediction through a random forest classifier, and sequentially undergoes four core processing steps of temperature scaling calibration, Gaussian kernel chemical rule scoring, adaptive probability fusion and continuity rule correction, and finally outputs a phase identification result and confidence. The temperature scaling calibration adopts a temperature coefficient T>1 to reduce model overconfidence. The adaptive fusion dynamically adjusts a fusion weight according to a model entropy value. The continuity rule correction adopts a continuous function to perform smoothing processing on the boundaries of easily confused phases. The application significantly improves phase identification accuracy and confidence reliability, and realizes quantitative evaluation of result reliability.

This invention discloses a method, system, and device for identifying the working state of forging operations, relating to the field of industrial automation technology. The method includes: S1 acquiring multi-view images of the working area; S2 detecting core production components and auxiliary operation-related objects using a YOLOv11 target detection model; S3 generating multi-dimensional instantaneous state codes and mapping them to basic states; S4 constructing a state time series and analyzing the macroscopic working state using a hierarchical state machine; and S5 outputting state and statistical information. This invention supplements the YOLOv11 loss function, instantaneous state encoding, and state transition algorithm formulas, solving the problems of incomplete state definitions, missing transition state identification, and inability to finely identify auxiliary operations in existing technologies. It achieves accurate and robust identification of the entire forging production process, providing data support for lean production management, with low deployment costs and strong scalability.

The present application relates to the technical field of image recognition, and specifically discloses a vehicle part recognition method and system, an electronic device and a storage medium. The technical scheme of the present application generates an enhanced image dataset by acquiring multi-angle vehicle images and performing image enhancement processing, and divides the fine-tuning process of a pre-trained visual model into three stages to progressively unfreeze and update the top, middle and bottom layer parameters in sequence. Performance evaluation and threshold determination are introduced during stage conversion, solving the problems of excessive consumption of computing resources and excessively long training time caused by full parameter fine-tuning relying on large-scale datasets in the prior art, and the problems of insufficient recognition accuracy and generalization ability caused by the difficulty of pre-trained models to adapt to multi-angle complex backgrounds. The present application reduces the model training cost, shortens the model training cycle, and improves the vehicle part recognition accuracy and generalization ability.

The application discloses a target ranging and automatic luggage following method based on behavior feature recognition, which comprises the following steps: S1, collecting video sequence data of the environment in front of the automatic luggage; S2, inputting the behavior recognition sequence into an improved VideoMAE model, introducing a phase masking mechanism, and generating a behavior vector; S3, performing identity association processing by using a FastReID algorithm; S4, generating a trajectory sequence; S5, inputting the trajectory sequence into an improved Social-GAN algorithm, and generating a predicted trajectory; S6, calculating the distance parameter between the target and the automatic luggage, and generating a position parameter and a trend parameter; and S7, driving the automatic luggage to perform direction adjustment and speed control according to the moving instruction. The application enhances the behavior feature representation capability and the trajectory prediction stability, and improves the target ranging accuracy and the continuity and reliability of the automatic luggage following process.

The application discloses a kind of hybrid classical-quantum extreme weather identification method, system and equipment for power system, and relates to power system technical field.The present application is aimed at solving the problems of existing extreme weather identification sample scarcity, data imbalance, insufficient feature expression, lack of power adaptability label, etc.The present application includes extreme weather label making on original meteorological data, obtaining labeled meteorological data set;Construct a hybrid classical-quantum generative adversarial network based on variational quantum circuit, and use extreme weather samples for adversarial training to generate expanded extreme weather sample data;Construct a hybrid classical-quantum deep neural network classification model containing a classical feature extraction layer, a variational quantum circuit feature mapping layer and a classification output layer, and use labeled data and expanded samples for training;Input the meteorological data to be identified into the trained classification model, and output the extreme weather identification result.This technical solution effectively improves the accuracy of extreme weather identification under small sample.

This invention discloses a method for separating and identifying the authenticity of AI-generated content in distorted images, belonging to the fields of image forensics, AI-generated content detection, computer vision, and image credibility verification. This method addresses the problem that original metadata, content credentials, and generation parameters are easily lost or invalidated after images are transferred through platforms, screenshots, screen captures, or printed images. First, the method acquires the image to be detected and reads the image data, extracting propagation state features such as resolution ratio, compression marks, boundary regions, moiré patterns, paper texture, noise residuals, and file structure. Then, based on these propagation state features, the propagation state of the image to be detected is identified, and the corresponding detection branch is invoked to determine the effective content area, extracting content authenticity features and acquisition authenticity features respectively. Next, weights are assigned to various features according to the propagation state and dynamically fused to calculate the content authenticity risk result and acquisition authenticity judgment result. Finally, an authenticity separation result is generated, outputting a detection report containing the propagation state, acquisition authenticity, content authenticity, AI generation risk level, evidence items, and uncertainty explanation. This invention can distinguish between the authenticity of the image acquisition method and whether the content carried by the image has the risk of AI generation, even when the original image file information is invalid. It is applicable to the auxiliary identification of AI-generated content in scenarios such as screenshots, platform transfers, screen re-photographs, printed re-photographs, and mixed transmission distortion.

The application provides a kind of unbalanced data classification model training method based on contrast learning, comprising: double view enhancement is carried out to training sample, and the training sample after enhancement is input into feature extraction network to obtain backbone feature;The backbone feature is input into projection head and classification head respectively, and the embedding vector in contrast learning space is used to update the prototype feature vector of each category at present, and the prototype feature vector of each category under the current batch is output;A joint loss function including class-aware contrastive loss function, class-balanced classification loss function and prototype consistency loss function is constructed;Feature extraction network parameters, projection head parameters and classification head parameters are updated simultaneously using back propagation algorithm.Class-aware contrastive loss function and prototype feature vector update enable tail classes to obtain stable intra-class aggregation constraints when same-class positive samples are scarce, reduce the dominant effect of head classes on feature space, and thus alleviate classification decision bias.

This application discloses a visual recognition method and system for violations by workers at substation construction sites. The method includes: acquiring a real-time video image sequence; identifying a set of spatial structural entities and key human node information based on the real-time video image sequence; extracting equipment image information and constructing a ceramic skirt edge contour model; determining spatial topological constraints by combining the set of spatial structural entities and the ceramic skirt edge contour model; calculating foot kinematic vectors; constructing a spatially accessible region by combining key human node information and spatial topological constraints; determining whether the worker's key foot nodes have entered an obstructed area; predicting and determining foot proxy nodes for obstructed areas; validating the foot proxy nodes and determining valid foot proxy nodes; and performing topological analysis of human force based on valid foot proxy nodes and key human node information to determine violations. This application can effectively improve the detection efficiency of violations by workers at substation construction sites.

The invention relates to the technical field of plastic scintillators, in particular to a plastic scintillator anti-coincidence detection method and system, and the method comprises the following steps: obtaining scintillator array electric pulse signals, recognizing a continuous triggering state, analyzing a triggering sequence through combining the space coordinates of a detector, and comparing center and peripheral triggering time intervals. And analyzing a center and peripheral trigger time relationship. According to the method, a continuous time sequence is constructed by reading trigger time and detector numbers, abnormal trigger signals are screened out in combination with pulse amplitude, so that detection data maintain stable trigger characteristics, and a particle trigger path is formed by analyzing adjacent trigger intervals and analyzing adjacent position direction changes according to array detector space coordinates and a trigger sequence. And a central trigger event chain is identified through the interval relationship between the central trigger time and the peripheral trigger time, so that anti-coincidence event judgment is established on the basis of time correlation and spatial path collaborative analysis, random trigger interference is reduced, and the particle event identification stability is improved.

The invention provides a photoetching defect seed physical constraint identification method and system, and the method comprises the steps: receiving the original data of a photoresist surface topology deviation field, carrying out the spatial normalization, and calculating the first-order derivative gradients in x and y directions; on the basis of a photoresist scattering model, marking a noise region in combination with the gradient amplitude and the scattering angle theta, and generating noise marking data; integrating the normalized deviation field, the x / y gradient and the noise mark as multi-channel data, and inputting the multi-channel data into a physical constraint CNN module; when the features are extracted, the dynamic weight alpha is calculated according to the photoresist viscosity eta, and the channel feature weight is adjusted; determining a coefficient beta by combining photoresist parameters, and calculating physical constraint loss function constraint model training; outputting a defect probability graph, carrying out threshold segmentation to obtain candidate seeds, and judging real defect seeds according to a curvature continuity index C and a closure degree; and outputting a real seed to the adaptive control unit, and feeding back noise features to perform online fine tuning on the CNN module.

This invention discloses a river channel crack point identification system based on simulated data and a deep learning architecture. The system includes generating a simulated dataset based on a river channel power erosion model, injecting random Gaussian noise into the elevation data, and performing hard negative sample mining on the steady-state samples of the simulated dataset. A cumulative minimum function is used to correct the physical consistency of the elevation data, and source shielding is applied to specific river channel data to obtain enhanced sample data. A three-channel physical signal matrix is ​​constructed based on the number of enhanced samples, and this matrix is ​​input into a deep residual feature extraction network to obtain crack point classification probabilities. The deep residual feature extraction network is then trained using regularization. The system performs a transfer from simulation to reality based on the deep residual feature extraction network, inputting a real river channel longitudinal profile and performing sliding window prediction to output the river channel crack point identification result. The river channel crack point identification result includes the probability of crack point existence and the crack point type.

The present application relates to a kind of underwater three-dimensional identification method for ship body sacrificial anode weld, it solves the technical problem that existing weld identification method for ship body sacrificial anode is difficult to accurately identify, three-dimensional reconstruction precision is low;It obtains multiple binocular images by binocular camera in different view angles and gathers the image of the region where underwater sacrificial anode is located, carries out pre-processing to multiple binocular images respectively, inputs the binocular image after pre-processing to instance segmentation model, instance segmentation model outputs the binary mask of weld, generates weld point cloud based on mask constraint, adopts region growing algorithm to separate out the ridge line of weld point cloud, obtains multiple ridge line point cloud clusters.The present application is applicable to ship cathodic protection system installation detection, underwater maintenance and automatic repair welding operation.

This invention discloses a speaker recognition method, system, storage medium, and device based on ordinary pronunciation. The method includes: acquiring real-time audio data and extracting spectral features based on the real-time audio data to obtain spectral features corresponding to the real-time audio data; inputting the spectral features corresponding to the real-time audio data into a trained UNET network to generate a spectral mask corresponding to the real-time audio data, and detecting whether the real-time audio data is an ordinary pronunciation based on the spectral mask; if the real-time audio data is an ordinary pronunciation, fusing the spectral mask and spectral features to obtain an enhanced spectrum corresponding to the real-time audio data; inputting the enhanced spectrum corresponding to the real-time audio data into a trained speaker embedding layer network to obtain a real-time speaker embedding layer corresponding to the real-time audio data; and comparing the real-time speaker embedding layer with the registered speaker embedding layer to identify the speaker corresponding to the real-time audio data.

The application discloses a music scene recognition method and system based on artificial intelligence, and particularly relates to the field of audio signal processing, wherein frequency, volume, rhythm, pitch and duration and other features are extracted by performing time-frequency analysis and filtering processing on input audio signals to construct an audio feature matrix; then the features are weighted in combination with scene differences to generate a weighted feature vector which is input into a deep learning model for recognition. Through model fusion and feedback optimization mechanism, the output result is dynamically adjusted, and the fuzzy label is secondarily confirmed, so that a stable and accurate scene label is finally output. The application realizes fine description of complex music signals and accurate scene recognition, effectively avoids the deficiency that traditional methods are easy to confuse in multi-style transition and fuzzy scenes, improves the ability of the system in label self-correction and result consistency, makes the recognition result have higher accuracy, robustness and practicality, and can be widely applied to intelligent music recommendation and multi-scene audio processing.

The application provides an ancient building fire identification method based on component occlusion sample images, and relates to the technical field of fire image identification.The ancient building fire image sample is obtained first, and the ancient building component occlusion template is constructed according to the component occlusion mode in the ancient building scene; the ancient building fire image sample is subjected to structure occlusion enhancement processing based on the ancient building component occlusion template, and the component occlusion sample image is obtained; the ancient building background image is obtained and the negative sample image is constructed; the ancient building fire image sample, the component occlusion sample image and the negative sample image are labeled, and the ancient building fire identification training sample set is generated; the fire identification model is trained by using the training sample set and the fire identification is carried out.The application can solve the problems of missing detection and false alarm of general algorithm models for occlusion detection in the ancient building scene and low recognition rate.

This invention relates to the field of environmental parameter detection technology, specifically to a self-driven temperature and humidity detection system based on LED geometric positioning. It solves the problems of high power supply dependence, large computational load for automatic recognition in pointer-type instruments, poor anti-interference capability, and difficulty in coordinating low-power recognition and self-driven power supply in existing temperature and humidity detection devices. The system includes a temperature detection component and a humidity detection component with pointers, a control component, an LED display component, an image acquisition component, a recognition processing component, a triboelectric nanogenerator component, and an energy management component. The LED display component includes a fixed reference LED group, and temperature LED groups and humidity LED groups that move synchronously with the pointer. This invention balances the needs of manual reading and automated recognition, significantly reduces the computational load and operating power consumption, improves recognition stability in complex environments, completely eliminates dependence on external power sources, and can be widely adapted to temperature and humidity detection needs in various scenarios.

The present application relates to the technical field of image processing, more particularly, the present application relates to a kind of based on machine vision's operating panel sheet printing positioning method, comprising: the image of operating panel sheet is collected;With any pixel point in the image as target point, the scarcity of target point in the image is calculated, and the target point corresponding to the maximum scarcity is selected as positioning point, real-time image positioning, the scarcity and the first scarcity eigenvalue and the second scarcity eigenvalue of the target point are positively correlated.The present application can effectively select the most unique positioning point in the image by introducing the concept of scarcity, combining the corner point distribution characteristics and the image feature frequency in the window, achieve high-precision positioning effect, through this method, repeated pattern interference can be avoided, and the recognition stability is improved, the accuracy of positioning can be significantly improved.

This invention discloses an automatic method and system for extracting landslide areas from remote sensing images based on few-sample enhancement and multimodal fusion. The method includes: acquiring and preprocessing multimodal remote sensing data; constructing a foreground-aware few-sample enhancement module, expanding the training samples through geometric transformation, spectral perturbation, and random cropping enhancement strategies; constructing a dual-branch feature extraction network to extract spectral and topographic features respectively; achieving adaptive fusion of spectral and topographic information through a cross-modal attention fusion unit; training a deep learning model using a composite loss function including cross-entropy loss, Dice loss, and boundary constraint loss; and performing topographic constraint post-processing and morphological optimization on the model output to obtain the automatic extraction result of landslide areas. This invention can achieve high-precision landslide identification under limited sample conditions, effectively reducing the false detection rate and improving the model's generalization ability and adaptability to complex scenes.

The application belongs to the technical field of image detection, and discloses a finite element cloud atlas crack fitting method based on direction prior and dark subinterval ROI, which comprises the following steps: calculating a preliminary direction angle of the crack according to the reference point coordinates, and taking the preliminary direction angle as the direction prior; taking the direction prior as the central axis direction, and taking one of the reference points as the benchmark to construct a strip type ROI; extracting the brightness values of the pixels in the strip type ROI, and screening the pixels with brightness values lower than or equal to a threshold value as a preliminary crack candidate point set; calculating the local gradient directions of the pixel points in the candidate point set, screening out the pixel points with a deviation between the local gradient direction and the direction prior less than a preset direction tolerance to form a direction consistency set; performing total least squares fitting on the pixel point coordinates in the direction consistency set to obtain a main direction straight line of the crack and fitting parameters, and superimposing the main direction straight line on the original finite element cloud atlas to generate a crack fitting result.

The invention relates to the technical field of image recognition and data processing, and particularly provides a customer information collection method and device based on OCR recognition, and the method comprises the following steps: S1, image collection and preprocessing; s2, performing OCR identification and character block extraction; s3, field title matching and content positioning; s4, field content extraction and multi-strategy verification are carried out; s5, data cleaning and format standardization; and S6, outputting a result and filling a form. Compared with the prior art, the method has high configurability and adaptability, and is suitable for various identity recognition and form filling application scenes.

The invention discloses a carrier speed estimation method and device based on a millimeter wave radar, and a storage medium, and relates to the technical field of carrier speed estimation. The method comprises the following steps: firstly, carrying out interference suppression processing on a distance-velocity spectrum matrix to obtain an interference suppression spectrum matrix; the first path performs velocity dimension energy aggregation on the interference suppression spectrum matrix to generate a static background energy distribution spectrum and determine a first velocity estimation value according to the static background energy distribution spectrum; a second path screens and verifies an effective ground reflection target from the interference suppression spectrum matrix, and a second speed estimation value is obtained by combining the pitch angle of the effective ground reflection target; and finally, respectively carrying out time sequence filtering and weighted fusion on the outputs of the two paths, and outputting the final carrier speed. According to the method, the problem that a traditional point cloud method fails in an open scene is fundamentally solved, and the robustness, accuracy and stability of a spectrum analysis method in a complex interference environment are remarkably improved.

The application relates to an oil-paper insulation device extended Debye model equivalent circuit modeling method, which comprises the following steps: acquiring measured data including a test frequency and a dielectric response current amplitude; establishing an extended Debye model equivalent circuit topology structure and a mathematical relationship between the dielectric response current amplitude and extended Debye model parameters; performing logarithmic transformation on the dielectric response current amplitude under different frequencies, and performing smoothing and sample expansion by adopting local weighted regression; constructing a parameter identification objective function, introducing a weight function to weight errors of different frequency points, balancing fitting weights and improving full-band fitting effects; adopting an artificial bee algorithm to optimize and identify the extended Debye model parameters, so that the objective function is minimum; and constructing the extended Debye model equivalent circuit according to the optimal parameter set, so as to realize extended Debye model modeling of the oil-paper insulation device. The application can improve the stability and accuracy of model parameter identification, and enhance the engineering applicability and repeatability of the modeling result.

The application relates to a multi-modal emotion recognition method and system fusing bidirectional attention and self-distillation, first, a bidirectional cross attention mechanism is designed to realize bidirectional interaction of information flow between modes, so that text, speech and video can be mutually corrected to mine deeper complementary clues. Secondly, a self-distillation type double-strategy learning framework is introduced to realize exploration of a global optimal solution through parallel conservative fusion and aggressive fusion paths, and the adaptability and generalization performance of the model in low-labeled data, noisy labeling and cross-domain migration scenarios are enhanced. Finally, a mode contribution balance mechanism is constructed to dynamically estimate the reliability and importance of different modes in different samples, suppress the interference of low-quality modes, and enhance the contribution of high-quality modes, thereby significantly improving the robustness and stability of the model in complex real scenarios. The application improves the fusion efficiency of text, speech and video modal information and the fine-grained emotion recognition precision.

The invention belongs to the technical field of communication analysis, and discloses an anti-crawler identification and flow control system and method based on behavior analysis. Comprising the following steps: analyzing at a data exchange network entry node to generate a session token; aggregating the browser static fingerprint, the transport layer protocol fingerprint and the interaction event feature sequence according to the session token to form a session feature frame; analyzing the consistency between the change amplitude of a negotiation field in the session feature frame and the change of a sequence element in the interaction event feature sequence to obtain a cross-layer association consistency degree; analyzing a gap change attribute in a retry confirmation event and a stability characteristic in a request interval sequence to obtain a replay and batch acquisition probability; and generating a control decision according to the cross-layer association consistency and the probability machine, packaging the control decision into a flow control instruction packet carrying a reason code and a session token, and issuing the flow control instruction packet to an execution node. According to the method, the stability of flow control in crawler recognition and jitter environments is improved.

The application provides a data construction and dynamic resampling fine-tuning method and system for multi-dialect speech recognition. Firstly, the existing speech recognition model is used to recognize and transcribe the dialect audio, clean it, and perform environment-related data enhancement on the audio based on acoustic environment simulation. Secondly, the dialect type, gender and speech speed features of the audio are extracted, and the data set is divided into multiple category buckets according to the feature combination. In the batch generation stage of model fine-tuning, exponential decay probability is used for sampling in the bucket to meet the allocated basic extraction quota, and dynamic balance scores are used to guide cross-bucket compensation. Finally, combined with the ladder data enhancement suitable for the number of historical sample extractions, the model parameter update is completed. The application solves the problem that the existing dialect recognition method cannot simultaneously consider multi-dimensional attribute balance, sample coverage and over-sampling risk control, and improves the recognition robustness and generalization ability of the model in harsh recording environments and multi-dialect cross-scenarios.

This invention relates to the field of emotional robots, specifically disclosing a voice emotion recognition system for an emotional companion robot. The system includes a voice input port embedded within the robot, a large-scale model processing system, and an emotion recognition output port. The voice input port collects voice data, which is then processed by the large-scale model processing system and output through the emotion recognition output port. The large-scale model processing system comprises a voice preprocessing module, a feature extraction module, a feature fusion module, an emotion classification module, and a result feedback module. This system optimizes the voice preprocessing flow, improves anti-interference capabilities, extracts seven types of multi-dimensional features, and comprehensively characterizes emotional information. It designs a 1D CNN network with "3 convolutional layers + 3 fully connected layers" to enhance global feature integration capabilities. It employs a triple data augmentation strategy and an optimized training strategy to improve model generalization ability and training effect. A result feedback module is added to optimize the recognition results in real time.