43results about How to "Fast reasoning" patented technology

The application provides a large model security management method, system, device and medium based on a core particle architecture, which can be applied to the technical field of semiconductor integrated circuits. The method comprises the following steps: setting a physically isolated special security core particle as a hardware trusted root in a heterogeneous integrated system, and connecting the special security core particle with at least one target core particle running a large model through a bus; deploying a lightweight security supervision model with a smaller parameter quantity than the large model in the special security core particle; collecting behavior characteristic data of the target core particle running the large model in real time; performing inference by using the lightweight security supervision model, calculating an abnormality metric between the current behavior of the target core particle and a preset normal behavior model; updating a dynamic trust state of the target core particle based on the abnormality metric; and performing a hardware-level security response operation by the special security core particle in response to the dynamic trust state reaching a predetermined threshold.

The present application relates to the field of digital image processing and information security, and discloses a document tampering detection method and system based on image data processing, comprising the following steps: first, extracting the noise residual and microscopic penetration characteristics of the document image, and constructing a physical potential energy field and a virtual viscous resistance field; then, using a Darcy law variant model for dynamic evolution, generating a virtual flow velocity vector field to simulate the sliding behavior of fluid in heterogeneous media; subsequently, constructing a heterogeneous graph based on the flow field divergence singular point and streamline trajectory, using a graph neural network to aggregate the node dynamics characteristics for deep reasoning, and finally generating a tampering positioning mask. The present application innovatively introduces fluid mechanics field theory, converts hidden static texture differences into significant dynamic flow field anomalies, solves the problem that the prior art is difficult to capture microscopic tampering traces, and significantly improves the detection accuracy and generalization ability in complex document scenarios.

The application discloses a kind of textile flaw detection methods based on machine vision, and the textile RGB image is obtained by image acquisition and preprocessing, and size adjustment, normalization operation is carried out, core detection process includes: feature extraction is carried out using ResNet50 network, and abnormal area recognition is realized in pixel level in combination with SPADE algorithm and k nearest neighbor, on this basis, using transfer learning strategy, specific textile flaw data set is fine-tuned and optimized, to build high-precision detection model, CFLOW-AD model is used for multi-scale feature extraction and the generation of abnormal score chart, accurately identify potential flaws, and integrated real-time feedback mechanism, system according to preliminary detection result, dynamically adjust detection parameter, and suspected area is re-detected, through TP, FP, ROC curve index carries out model evaluation, generates detailed flaw report, the method effectively improves the automation level and accuracy of textile flaw detection.

The application discloses a dialect self-adaptive learning method and system, comprising a structured dialect domestication interface, a speech intent recognition unit, a grammar framework domestication unit, a cross-dialect analogy reasoning engine, a dialect-scene-customs linkage memory module and a distributed dialect knowledge federation network; the robot can continuously update dialect recognition capability through daily interaction without intervention of manufacturers. Users can teach the robot to learn dialects in a natural way such as demonstration, error correction and praise, just like teaching children to speak. The system can continuously optimize the model according to the feedback of users, and adapt to the language habits and dialect variants of different users.

The application discloses a kind of based on improved YOLOv10m light weight small target enhancement detection model noodle adhesion identification method, belong to food industrialization intelligent production and machine vision target detection cross technical field. Including: the original image of acquisition noodle industrialization cutting link;Establish noodle adhesion state detection model and training;Using the model detects noodle adhesion state.The noodle adhesion state detection model of the application, with YOLOv10m as baseline model, adopts light weight improved NAD-MobileViT main network, to adapt to the extraction demand of noodle slender structure characteristics;Introduce CEECA module in Neck feature fusion layer, realize the channel-space joint reinforcement of core feature and redundancy information suppression;In Head detection head part, add new high-resolution small target detection head.It can realize the real-time accurate monitoring of noodle edge form after cutting, can effectively identify interlaced state.

The application provides a method and system for image data recognition model construction. The method comprises: obtaining an image data set comprising true value labels; based on the image data set and a teacher network model, using an FSP method to initialize the weights of each layer of a student network to obtain at least two student network models with the same network structure; based on the image data set comprising true value labels, using a DML distillation method to train each student network model, and when the trained student network models all meet preset conditions, taking the trained student network models as an image data recognition model. Through the method, at least two student network models with the same structure and completed weight initialization of each layer are obtained by transferring learning from a teacher network model, which can avoid problems such as the inability to well initialize a neural network due to insufficient data, and can reduce the training difficulty. The DML distillation method is used to obtain an image data recognition model with guaranteed recognition ability, and the training cycle is short, the speed is fast, the deployment is easy, and the applicability is wide.

The application relates to an image super-resolution reconstruction method and device based on anchor point guidance and boundary refinement cooperation, and belongs to the field of computer vision. The method comprises the following steps: performing shallow feature extraction on an input low-resolution image to obtain initial features; a deep feature enhancement network comprising multiple series cooperative reconstruction units is constructed, the initial features are input, and the initial features are sequentially enhanced by each unit; each unit performs anchor point routing global interaction and boundary perception region refinement on the input features to obtain global enhanced features and local refined features, the global enhanced features and the local refined features are fused through adaptive gating, and then the input features are fused through residual fusion, current cooperative enhanced features are output and are transmitted to the next unit, and finally, deep enhanced features are obtained; upsampling and pixel reconstruction are performed on the deep enhanced features, and a super-resolution image is output. The application considers long-distance structure dependence modeling and local boundary detail recovery, and realizes the cooperative optimization of global structure consistency and local detail fidelity under the constraint of lightweight deployment.

The present application relates to a ship heave motion prediction method and system based on a lightweight context-aware network, belonging to the field of ship and ocean engineering motion control technology, including the following steps: signal acquisition and preprocessing, heave motion multi-step prediction based on LCGNet network, model training and optimization, quantization and compression of the trained LCGNet model, and deployment on a shipborne edge computing device; during system operation, historical heave data is collected in real time and input into the model, and step S2 is executed in a loop to realize continuous online prediction of future heave motion; through innovative lightweight network structure design, high-precision, multi-step heave motion real-time prediction is realized at extremely low parameter and calculation cost to meet the engineering deployment requirements of the shipborne active heave compensation system.

The application belongs to the technical field of natural language processing and information extraction, and specifically relates to a fine-grained entity extraction method and system for complex documents, which comprises the following steps: obtaining an OCR text box set of a document image, extracting text content and coordinate information of each text box, and fusing the text features and visual features to generate a multi-modal fusion feature matrix; calculating geometric position relationships according to the coordinate information of each text box, modeling and fusing spatial correlations by using a multi-scale kernel function to generate a spatial bias matrix; dividing all attention heads into geometric constraint heads and global free heads, calculating attention and fusing after multi-head linear projection to obtain context-enhanced features; and calculating label emission scores, combining sequence constraint values to optimize model parameters, and outputting entity extraction results through global optimal sequence search. The application does not require external pre-training data, has small parameter quantity and fast reasoning speed, and can be used for entity extraction of complex documents such as government forms and commercial bills.

This invention discloses a fast activity recognition method based on a multi-path parallel MLP mixer architecture. First, multi-dimensional sequence data is used as input. An embedding module encodes the multi-dimensional data to obtain a feature sequence. This feature sequence is then fed into multiple MLP mixer branches, where features are mixed along the time, channel, and frequency domain dimensions, respectively. In each branch, after passing through multiple MLP mixers, global pooling is used to aggregate the feature sequence along the time dimension into a feature vector. Finally, the feature vectors obtained from each branch are concatenated and sent to a classification head for activity recognition. Compared with existing deep learning-based activity recognition methods, this invention achieves more accurate activity recognition with fewer model parameters, less floating-point computation, and faster inference speed, meeting the requirements of high precision and high efficiency in practical applications.

The application provides a human action recognition and quality evaluation method and system based on multi-modal data fusion, the application collects RGB visual data, 3D skeleton data, inertial IMU data and electromyography sEMG data of human action, and carries out corresponding pretreatment; through branch parallel feature extraction, RGB visual feature vectors, 3D skeleton feature vectors, IMU inertial feature vectors and sEMG electromyography feature vectors are obtained respectively; based on the modal reliability values of each branch, multi-modal fusion feature vectors are obtained by dynamically distributing and fusing the weights of the cross-modal attention mechanism to the multi-path features; the action recognition result is obtained by using a lightweight action recognition network; and the action recognition result is quality evaluated through the comprehensive action quality score. The application can not only accurately recognize the action category, but also quantize the action quality from three dimensions of form, dynamics and force, thereby improving the robustness and accuracy of action recognition in a complex scene.

This invention discloses a small target detection method and system based on improved YOLOv11 and fused attention, applicable to small target detection in unmanned aerial vehicles (UAVs). The method includes the following steps: constructing a backbone network using a custom module SFBCR; constructing a neck network using a custom module MFBCR to obtain a custom model; and designing a fused attention mechanism based on the SimAM attention module and PartialConv partial convolutional modules. The model parameters are determined by training on the public UAV small target dataset VisDrone. Compared with existing neural networks, the proposed detection method achieves an average accuracy improvement of 2.03% in UAV small target detection applications while maintaining detection speed. It can be used in smart city construction to improve urban traffic management.

The invention relates to the technical field of computer simulation, and particularly provides a material pile form change prediction method, which comprises the following steps: firstly, establishing a simulation system of a to-be-operated material pile, obtaining a material pile height map before spading and a corresponding bucket posture, and inputting the material pile height map and the corresponding bucket posture into a pre-trained double-agent cooperation model to predict the height of the material pile after spading; the double-agent cooperation model comprises a first agent and a second agent, and the first agent inputs a material pile height map and a bucket attitude before spading and predicts and outputs a global probability mask; and the second intelligent body inputs a material pile height map before spading, a bucket attitude and a global probability mask, predicts and outputs a global height change map, and further generates a material pile height map after spading to update a simulation material pile. According to the method, a double-agent collaborative architecture model is adopted to replace a traditional DEM method, and the calculation amount is greatly reduced while the result accuracy is guaranteed.

The application provides a kind of multi-modal pre-training model training method and device, constructs the multi-modal pre-training model containing multimodal graph-text information processing network;Weak alignment image-text data set is constructed;Wherein, weak alignment image-text data set contains text data set, image-label data set and image-reference description data set;Multi-modal pre-training model is trained using weak alignment image-text data set.The multimodal graph-text information processing network of the application can directly process multi-modal graph-text information, without external model auxiliary extraction image feature, and has strong inference ability.At the same time, weak alignment image-text data set is used to train multi-modal pre-training model, reduce the dependence on artificial labeling image-text alignment data, avoid the problem of large data overhead that occurs when using alignment large-scale image-text data set to train multi-modal pre-training model.

This invention relates to a three-dimensional defect segmentation method, system, and storage medium, belonging to the interdisciplinary fields of computer vision, deep learning, and nondestructive testing. The three-dimensional defect segmentation result is obtained by inputting the industrial CT volume data of the object to be inspected into a 2.5D segmentation network composed of an encoder, a spatial feature displacement module, and a decoder. This invention achieves efficient three-dimensional context awareness, significantly improving defect segmentation accuracy and three-dimensional morphological integrity. The segmentation result has topological connectivity and smooth boundaries, and its computational complexity and memory usage are far lower than those of a full 3D network. It can meet the real-time requirements of industrial online inspection and has broad application prospects in high-end manufacturing.

This invention belongs to the field of computer vision technology, specifically relating to a skin lesion image segmentation method based on a Transformer dual-branch model. The method constructs and trains a Transformer dual-branch model, inputting the image to be processed into the trained Transformer dual-branch model to obtain the segmentation result. The Transformer dual-branch model includes a main branch network, an auxiliary branch network, and an information aggregation module. This invention proposes a novel skin lesion image segmentation method that addresses the shortcomings of traditional deep learning methods in extracting global contextual information. It utilizes an efficient multi-scale visual Transformer as an encoder to extract more powerful and robust features. Simultaneously, it introduces low-level feature modules and high-level feature fusion modules to effectively improve the network's feature learning ability and segmentation performance.

The invention discloses a fetal distress auxiliary diagnosis system based on continuous wavelet transform and a shallow convolutional neural network, and the system comprises a data obtaining and preprocessing module which is responsible for obtaining an original fetal heart rate signal and carrying out the data preprocessing of the original fetal heart rate signal; the time-frequency characteristic graph construction module is responsible for converting the signals processed by the data acquisition and preprocessing module from one-dimensional data into a two-dimensional time-frequency characteristic graph by using continuous wavelet transform; and the classification and recognition module is responsible for inputting the time-frequency characteristic pattern into a shallow convolutional neural network model and outputting a fetal distress state classification result. According to the method, a one-dimensional time sequence signal is converted into a two-dimensional time-frequency image, and joint distribution information of the signal in a time domain and a frequency domain is subjected to visual coding, so that a shallow convolutional neural network model can fully utilize the advantages of the shallow convolutional neural network model in image processing, deep pathological features hidden in an FHR signal are mined, and the deep pathological features hidden in the FHR signal are extracted. And the defect that the one-dimensional convolution is insensitive to the time sequence is overcome.

This invention provides a safety detection method for the hook of an overhead crane used for transporting molten metal in a steel plant, belonging to the metallurgical field. The method includes: S101, acquiring visual signals of the hook engagement or disengagement process between the hook and the ladle trunnion; S102, acquiring the crane weight signal and comparing the current weight signal with the average of the previous three weight signals. If the current weight signal is greater than the average, it indicates that the ladle has begun engagement, proceeding to S103; if the current weight signal is less than the average, it indicates that the ladle has begun disengagement, proceeding to S103; otherwise, returning to S101; S103, using a cloud-edge system architecture, the equipment in the visual signal processing cabinet processes the acquired visual signals through a deployed model to determine whether the engagement or disengagement process is safe and whether the ladle's posture after hoisting is safe; S104, if unsafe, the equipment in the visual signal processing cabinet issues an alarm signal. Using this invention, the safety of the engagement or disengagement process can be checked in a timely manner, and an alarm can be triggered promptly.

The invention provides a WOA-GRU-based hydrogen atomic clock fixed parameter control method and device. The method comprises the following steps: acquiring and preprocessing a historical operation parameter time sequence of a hydrogen atomic clock; constructing a parameter time sequence prediction model taking a gating circulation unit as a core; performing global automatic optimization on the hidden layer size, the discard rate and the learning rate of the GRU network by adopting a whale optimization algorithm; training a GRU model by using the optimized hyper-parameters; and finally, obtaining a group of collaborative optimal fixed parameter values capable of enabling the hydrogen atomic clock to stably operate for a long time by utilizing the trained model in a rolling prediction mode. The GRU network with a more concise structure is adopted to replace a traditional LSTM network, intelligent global search of WOA is combined, the training efficiency and the model stability are remarkably improved while the prediction precision is guaranteed, and the method is particularly suitable for a hydrogen atomic clock control scene with limited data volume and high parameter noise and has good application prospects. Accurate, stable and intelligent fixed control of internal parameters of the hydrogen atomic clock is realized.

The application discloses a network security management and control method and system based on different scenes, and the method comprises the following steps: when the number of elements in a to-be-responded queue is less than a preset number, performing SQL injection detection on each request parameter based on a first interception rule to obtain an SQL injection detection result; when the number of elements in the to-be-responded queue reaches the preset number, simultaneously performing SQL injection detection on each request parameter based on a second interception rule to obtain an SQL injection detection result; and based on the SQL injection detection result obtained based on the first interception rule or the second interception rule and the request parameter to be responded, performing a corresponding response operation. The application can provide corresponding SQL injection detection schemes for different scenes, and can accelerate the response speed of a large number of SQL access requests in a short time.

The invention relates to the technical field of aerospace equipment detection, and particularly discloses an engine turbine blade defect detection method and system based on deep learning. Blade static images, dynamic videos and real-time image data are collected through an industrial camera, a video recorder and an external camera; after the image is subjected to denoising, normalization and enhancement preprocessing, an improved YOLOv8 deep learning algorithm is utilized to construct a defect detection model, and four types of defects including scratches, oil stains, rust stains and damage are accurately recognized; establishing a defect evaluation model based on multi-dimensional indexes such as an accuracy rate, a recall rate, F1-Score, mAP and the like, dividing the defect severity into four levels of slight, moderate, serious and fatal, and generating a detection report containing visual annotation and processing suggestions; and results are fed back to related departments in real time through the display screen, the data interface and the mobile terminal. The method is high in detection precision, high in efficiency and high in scene adaptability, a scientific basis is provided for blade manufacturing quality control and maintenance decision making, and operation safety of an aero-engine is effectively guaranteed.

The present disclosure relates to the technical field of image processing, and particularly relates to a method for determining an image restoration model, an image restoration method and device, which comprises: acquiring a reference multi-focus image and an initial model, wherein the reference multi-focus image comprises multiple reference images with different focus positions; determining full-focus label data corresponding to the reference multi-focus image according to the multiple reference images with different focus positions; determining an input image by using a point spread function according to the full-focus label data; and training the initial model based on the input image and the full-focus label data to obtain an image restoration model. The technical scheme of the embodiment of the present disclosure improves the image restoration quality and realism, and accelerates the restoration speed.

The application relates to the field of ocean monitoring and forecasting, and discloses a buoy trajectory prediction method based on stationary wavelet and wind-flow-position interaction, which comprises the following steps: acquiring a trajectory sequence of a buoy in a historical period and corresponding ocean environment field data and performing standardization processing; decomposing the normalized trajectory sequence of the buoy through stationary wavelet transformation to obtain a plurality of sub-sequences of different scales, and then obtaining a plurality of segmented token sequences through dynamic segmentation; performing feature extraction on each segmented token sequence after feature mapping through a wind-flow-position interaction feature extraction mechanism; reconstructing the interaction features output by the wind-flow-position interaction feature extraction mechanism through stationary wavelet inverse transformation, and finally outputting a trajectory coordinate prediction value of the buoy in a future period after linear layer mapping. Through the introduction of wavelet multi-scale decomposition and dynamic segmentation, the model can explicitly capture the trend and details in the trajectory, and effectively cope with non-stationarity.

The invention relates to the technical field of artificial intelligence, in particular to an AI-based intelligent image processing system and method.The AI-based intelligent image processing system comprises an image input interface, an image analysis module, multi-dimensional features of images, a dynamic task routing pool, a lightweight processing module chain and an image output interface, and the image input interface is used for obtaining original images. According to the AI-based intelligent image processing system and method, through the front-end analysis module, the processing flow and intensity are dynamically determined according to the characteristics of the image, and excessive deletion and optimization are avoided; through a dynamic routing mechanism, the overall efficiency and effect are improved; and meanwhile, each sub-module adopts lightweight designs such as depth separable convolution and channel attention, the overall model is small in parameter quantity, high in reasoning speed and easy to deploy in a mobile terminal or embedded equipment, a modular design and a dynamic combination strategy enable the system to better adapt to images of different sources and different degradation types, and the robustness and generalization ability are higher.

This invention provides a method for accelerating the compression of intelligent models on edge terminals, relating to the field of edge intelligence. Its key feature is that it includes: acquiring real-time operating status data of the edge terminal and feature information of the current input data, wherein the real-time operating status data includes computing resource load, memory occupancy, battery level, and device temperature. The advantages of this invention are: by sensing the device status and input complexity in real time, and utilizing reinforcement learning to dynamically generate an optimal strategy combination including pruning, hybrid quantization, and early termination mechanisms, this method overcomes the limitations of traditional static compression, achieving a dynamic balance between computing power, accuracy, and energy consumption. It not only significantly reduces inference latency and memory usage but also adaptively adjusts based on battery level and temperature, effectively solving the problems of difficult deployment, high heat generation, and short battery life of large models on resource-constrained edge devices, and greatly improving the real-time performance and stability of intelligent applications.

The application discloses a personal image privacy protection method, system and electronic equipment. The method provided by the application allows a user to specify a certain person in a protected image through language expression, generates multi-scale visual features with sufficient fusion of image and text information by using a lightweight deep neural network to perform parallel coding on input reference information and personal images, generates a stable specified personal privacy protection image mask through a multi-scale feature fusion and mask positioning enhancement module in the decoding process, and in addition, introduces a balanced binary cross-entropy loss to solve the pixel imbalance problem in training, optimize the network performance, and improve the personal image privacy protection effect. The application can solve the problems of excessive protection of content and pixel imbalance in the training of the reference personal image privacy protection network in the existing personal image privacy protection technology.

The application discloses an intelligent triage method and system based on pre-diagnosis text clustering and pattern recognition, and the method comprises an offline training stage and a real-time triage stage. In the offline training stage, a large amount of historical pre-diagnosis texts are acquired and preprocessed, and after being vectorized by a pre-trained language model, the semantically similar vectors are aggregated by using an unsupervised clustering algorithm, the clustering cluster-department mapping is completed in combination with a 'department-symptom' medical knowledge base to generate a pseudo-label data set, and then an intelligent triage model is trained; in the real-time triage stage, after being preprocessed and vectorized, the complaint text of a new patient is input into the intelligent triage model to output a department prediction probability and recommend an optimal department. The method does not need large-scale artificial labeling data, greatly reduces the training cost, and simultaneously improves the triage accuracy and real-time response speed.

The invention provides a deep counterfeit content detection system and method based on a multi-modal large model, and the method comprises the steps: taking a video clip as input, employing a phased Video Swin Transform as a backbone network, and extracting multi-scale deep space-time visual features in video data; a Mel-frequency cepstrum coefficient fragment is used as input, a cascaded ERes2Net module is used for constructing an audio feature extraction network, a multi-stage fusion mechanism is introduced, extracted audio features are injected into a visual feature extraction branch stage by stage, and fused features are formed; according to the method, features are aggregated through parallel space-time pooling and a time attention mechanism, then internal association of audio-visual features is deeply mined in a local region, among all regions and in three dimensions of local-global through a core local-global interaction module, and finally, a deep forging detection result is output through a full connection layer. According to the scheme, the optimal effect is achieved on the largest deep counterfeiting detection data set in the industry.

This application relates to the fields of artificial intelligence and digital healthcare, and provides a target detection method and related equipment based on pre-trained networks. The method includes: acquiring a trained first target detection model, which includes multiple pre-trained network layers; then quantizing each pre-trained network layer of the first target detection model to reduce the model parameters from initial accuracy to target accuracy; using the quantized first target detection model as a second target detection model; acquiring a target image; inputting the target image into the second target detection model to obtain a prediction result for the target image. The embodiments of this application can improve model inference speed and reduce hardware resource consumption while ensuring target detection accuracy.

This invention discloses a method, system, and device for estimating the area of ​​the eye muscles and the backfat thickness of live pigs based on deep learning. The method includes: acquiring and constructing a dataset of live pig ultrasound images; selecting regions of interest (ROIs) in the ultrasound images, preprocessing them, and standardizing the input size; inputting the preprocessed ultrasound images into a ReAMS-UNet neural network for semantic segmentation of the eye muscle region; filtering the image contours to remove false positives and false negatives in the segmentation results; determining the upper and lower boundaries of the backfat thickness through image binarization; calculating the eye muscle area and backfat thickness, and outputting the calculation results. Based on a large-scale, highly diverse dataset of pig ultrasound images, this invention utilizes ReAMS-UNet to integrate residual learning for stable training, a hybrid attention mechanism for adaptive feature optimization, multi-scale fusion for contextual and spatial information fusion, and auxiliary supervision for enhanced gradient propagation. It achieves high segmentation accuracy, fast inference speed, accurate trait estimation, and results that reflect true carcass traits. The process is automated and suitable for high-throughput analysis applications.