40results about How to "Reduce inference time" patented technology

The invention discloses a task processing method and device based on a large language model, equipment and a medium, and the method comprises the steps: carrying out the disassembly of a target task through the large language model, so as to obtain a plurality of subtasks and subtask information corresponding to the subtasks; based on the task description, the task category and the task difficulty corresponding to the sub-tasks, determining execution models corresponding to the sub-tasks; executing the plurality of sub-tasks by using an execution model to obtain a plurality of sub-execution results; and integrating the plurality of sub-execution results to obtain an execution result of the target task. The category and difficulty of an input task are identified by using a large model, and a complex task is disassembled into a plurality of sub-tasks. According to the difficulty level of each subtask, determining an execution model which is most matched with the current task in difficulty and scene in a model pool; large models of different parameter volumes are dynamically called, and resource waste is reduced to the maximum extent while task quality is guaranteed.

The invention discloses a full-automatic rapid detection and segmentation method for photovoltaic cell defects, and the method employs a lightweight SSA-ViT image encoder to replace a heavy-weight ViT image encoder in SAM, employs a separable self-attention mechanism (SSA) to replace a multi-head attention mechanism (MHA), greatly reduces the complexity of the attention mechanism, reduces the calculation amount, and improves the detection efficiency. Compared with the prior art, the method reduces the reasoning time of a single-frame image from average 4099 ms to average 11.7 ms (350 times), greatly improves the detection efficiency of the model, takes the prediction result of the YOLOv10 model as a new segmentation normal form of YoMo-SAM prompt information, abandons the man-machine interaction segmentation normal form of the original SAM, and truly achieves the full-automatic target detection and segmentation task on the premise of guaranteeing the segmentation precision. Meanwhile, task generalization of zero samples is achieved, the problem of data shortage is solved while manual segmentation is avoided, and then the detection segmentation precision of the model is improved.

The invention discloses a real-time auricle identification and auricular point positioning method and system for multi-task feature sharing. An auricle identification and auricular point positioning combined model used in the method comprises an encoder and a decoder. The encoder includes a backbone network and a neck network. The decoder comprises a plurality of detection heads. A plurality of detection heads receive outputs of different layers of the neck network. The detection head comprises a target regression branch, a target classification branch and a feature point regression branch which are parallel. A plurality of detection heads share output features of the same encoder, and parallel target regression branches, target classification branches and feature point regression branches in the detection heads are respectively used for calculating loss functions of different tasks; therefore, auricle identification, frame selection and auricular point feature point marking tasks can be efficiently completed; on the basis, high-frame-rate real-time prediction is realized in the field of auricular point positioning for the first time, and the limitation of the prior art is broken through.

The invention discloses a joint detection and re-identification multi-target tracking method based on feature decoupling, and relates to the technical field of multi-target tracking. The objective of the invention is to solve the problem of low tracking precision caused by conflicts in optimization directions in a detection and re-identification task joint training process in an existing joint detection and re-identification multi-target tracking method. The method comprises the following steps: preprocessing a multi-target tracking data set; sending the processed image and label to a backbone network to extract a feature F; decoupling the feature F in the channel dimension and the spatial dimension in sequence to obtain a feature Fdet suitable for a detection task and a feature Fid suitable for a re-identification task; obtaining the position of a target detection bounding box in the image according to the Fdet subdivision detection branch; calculating a similarity matrix D according to the Fid and IOU overlapping degree; and matching the target detection bounding box in the image with the existing trajectory by using the D and a Hungary algorithm to finally obtain a tracking result. Experiments prove that when the method is used for multi-target tracking, the tracking performance can be improved.

The invention discloses a lake water body extraction method, system and device, a medium and a product, and relates to the field of target recognition, and the method comprises the steps: extracting a water body in a remote sensing image of a to-be-detected region based on a lake water body extraction model, obtaining a lake water body image, and determining a water body region of the to-be-detected region; the category of each pixel point is marked in the lake water body image, and the category of the pixel point is a water body or a non-water body; the lake water body extraction model comprises a lightweight feature extraction network, a global average pooling layer, an edge extraction module, a first convolution attention module, a second convolution attention module and a feature fusion module. According to the method, the complexity of the lake water body extraction model is reduced on the premise of ensuring the accuracy of the lake water body extraction model, the reasoning time of the lake water body extraction model can be reduced, the category of each pixel point in the remote sensing image is determined by adopting the lake water body extraction model, and the accuracy and efficiency of lake water body extraction are improved.

The invention provides a target detection method and device, equipment and a storage medium, and relates to the technical field of automatic driving visual perception, and the method comprises the steps: determining each lane line according to the original image data of a current lane, generating a plurality of first coordinate points of each lane line, and carrying out the sorting according to the first coordinate points, determining an area between the leftmost lane line and the rightmost lane line as a first maximum drivable area; projecting the first maximum drivable area and the plurality of first coordinate points of each lane line into the original image, and generating a second maximum drivable area in the original image and a plurality of second coordinate points of each lane line in the original image; and carrying out zero setting on an area outside the second maximum drivable area in the original image to obtain a pre-processed image, and executing target detection operation. According to the method, more key information needed by target detection is added, only the driving area is reserved in the original image, invalid information in the original image is reduced, and therefore the reasoning time of a deep learning model can be shortened.

The invention relates to an intelligent integrated circuit and an intelligent chip in an artificial intelligence hardware platform. The invention relates to the field of deep neural networks in an artificial intelligence general technology, in particular to a data deployment system, a training method, an inference method and a chip of a deep neural network. The deep neural network data deployment system comprises a storage module; the matrix decomposition module is used for decomposing the original matrix into the sum of N sub-matrixes, and N is greater than or equal to 2; wherein one of the sub-matrices is further decomposed into a product of two or more low-rank matrices; the matrix expression processing module is used for expressing the sub-matrixes which are not decomposed at first precision and then storing the sub-matrixes; expressing the two low-rank matrixes with second precision and then storing the two low-rank matrixes; wherein the original matrix is a weight matrix and / or an activation value matrix, and the first precision is lower than the second precision. According to the method, on the premise that the performance of the model is basically not lost, the storage space required in the training and reasoning process is effectively reduced, and the calculation efficiency is improved.

The application provides a large model inference acceleration method, device, equipment, storage medium and program product, and relates to the technical field of artificial intelligence, and the method comprises the following steps: inputting an interactive prompt word and historical dialogue into a large language model for self-recurrence decoding to obtain a current word element; it is detected whether the current word element is a text operation semantic word element; if the current word element is a text operation semantic word element, a target text segment is copied from existing text information based on the text operation semantic word element; a plurality of target word elements are updated to the end of the output text; based on the large language model, self-recurrence decoding is continued to generate a next word element, the next word element is taken as the current word element, and the step of detecting whether the current word element is a text operation semantic word element is returned to until the large language model generates an end symbol, and an updated output text is obtained. Through the above manner, the text copying efficiency is improved, the inference acceleration effect of the large language model is optimized, and the accuracy and stability of the copying operation are beneficially ensured.

The invention belongs to the field of computer vision and video generation, and particularly discloses a portrait animation generation method and system based on consistent distillation. The method comprises the steps that a reference image sample and a human body posture sequence sample are collected to serve as a training set, a pre-trained teacher model is constructed based on the training set, and the teacher model is used for generating a target video; initializing a consistency distillation model to obtain a student model, and transmitting knowledge of the teacher model to the student model through a track segmentation consistency distillation method; optimizing the student model through auxiliary loss optimization, dynamic region weighted optimization and facial detail injection to obtain an optimized model; and inputting a reference image to be reasoned and the human body posture sequence into the optimization model to obtain a portrait animation video. According to the invention, the time sequence consistency, coherence and detail fidelity of the generated video can be effectively improved.

The invention provides a target detection method and device combining a multi-branch decoupling head and feature enhancement. According to the method, a YOLOv5 model is selected as a reference framework to design a target detection network for target detection; the target detection network includes a backbone network, a neck network, and a head module. The method comprises the following steps: firstly, performing feature extraction on an input image by using a convolutional neural network through a backbone network to obtain three feature maps of different levels; the neck network performs feature enhancement on the three feature maps of different levels and then performs feature fusion so as to further extract features; the feature enhancement is added into cavity convolution processing to enlarge a receptive field; the head module adopts a three-branch decoupling head, independently divides category prediction, prediction frame positioning and confidence calculation in target detection into three network branches, performs convolution operation on extracted features, and outputs category prediction, prediction frame positioning and confidence calculation results. By using the method, the small-scale target detection precision of the remote sensing image can be improved.

This application relates to the field of data processing technology, and in particular to a method, apparatus, device, and storage medium for model training and 3D mesh reconstruction, used to rapidly generate high-quality 3D results. The model training method includes: obtaining an initial repair model by distilling a student model to be trained; the initial repair model is used to make the input initial sample set, after M iterative repairs, present the shape of a 3D point cloud corresponding to the initial sample set; obtaining an intermediate repair model by distilling the initial repair model; the intermediate repair model is used to make the input intermediate sample set, after N iterative repairs, present the shape of a 3D point cloud corresponding to the intermediate sample set; and obtaining a target repair model by adversarial training of the intermediate repair model based on the initial repair model, so that the error between the results obtained by repairing the same target sample set using the target repair model and the initial repair model conforms to a preset error range.

This invention discloses a lightweight joint estimation method for battery state of charge (SOC) in resource-constrained MCUs, comprising: extracting the weight matrices and bias terms of each layer of a trained LSTM time-series estimation network and flattening them into a one-dimensional static constant array; constructing a custom pure C language matrix operation library independent of a third-party AI inference framework on the target microcontroller, employing a global static memory pool management strategy, and defining a unified structure containing the number of matrix rows, columns, and one-dimensional floating-point data pointers at the C language level; the target microcontroller calling the custom pure C language matrix operation library to directly read the one-dimensional static constant array to complete the forward inference calculation of each gate unit of the LSTM network and output the prior SOC observation value; and calling the custom pure C language matrix operation library to perform state equation iteration of the Kalman filter and output the SOC estimation result. This invention enables the safe and efficient operation of complex coupled algorithms on a very low computing power platform.

The invention relates to a small sample point cloud semantic segmentation method and device, computer equipment and a storage medium. The method comprises the steps that support set point clouds and query set point clouds of corresponding masks are acquired, feature extraction is carried out on the support set point clouds and the query set point clouds, and support set features and query set features are obtained; according to the distance between the support set features and the query set features, aligning the support set features according to the distance to obtain the aligned support set features; obtaining a corresponding prototype according to the aligned mask supporting the set features, splicing the prototype with the query set features to obtain expanded query set features, extracting point cloud feature distribution information from the expanded query set features, and correcting the point cloud feature distribution information to obtain a point cloud feature distribution information set; and segmenting a corrected prototype and a corrected query set feature from the corrected point cloud feature distribution information. The method can reduce the reasoning time, and greatly reduces the point cloud segmentation error.

The invention relates to a lightweight behavior recognition method, and belongs to the technical field of behavior recognition. The method comprises the following steps: inputting collected video data and carrying out preprocessing; performing time domain segmentation on the preprocessed video data through a sparse sampling strategy, and extracting optical flow image features and an RGB image sequence; inputting the segmented and sampled optical flow image features and the RGB image sequence into a MobileNetV3 network for feature extraction, performing weighting processing on the features by using an SE attention mechanism, and lightening the whole network by using channel pruning; and fusing the features extracted in a plurality of time periods by adopting a segmentation consensus strategy, and finishing behavior recognition by combining classification results of time domain features and spatial features by using a maximum pooling method. According to the method, efficient behavior recognition is realized, and relatively high recognition accuracy can be achieved while relatively low calculation overhead is ensured.

The application provides a multi-scene safety inspection method and system based on a UAV, and belongs to the technical field of UAV safety inspection, and comprises the following steps: constructing a multi-scene safety sample library; training a multi-task computer vision model based on the multi-scene safety sample library to obtain a safety inspection algorithm model integrating multiple recognition capabilities; the multi-task computer vision model comprises a shared feature extraction backbone network and multiple independent task head networks; a training set is input into the shared feature extraction backbone network to extract general feature maps; the general feature maps are respectively input into the multiple independent task head networks for parallel training to obtain the safety inspection algorithm model; and the safety inspection algorithm model is loaded on a UAV flight platform to perform parallel inference analysis on inspection images collected by the UAV in real time, and to identify preset safety hidden danger targets.

The invention discloses a crack detection method and system based on edge end equipment, and the method comprises the following steps: collecting an image frame of a crack condition to be detected, and carrying out the preprocessing of the image frame; performing matrix construction on the preprocessed data, and performing image feature extraction on the preprocessed data by using convolution calculation; performing quantization processing on a feature extraction result by using an INT8 quantization algorithm; inputting the quantized data into a Leaky ReLU function to carry out data nonlinear transformation; constructing a 2X2 data matrix, and processing the data after nonlinear transformation by using a maximum pooling method; carrying out fusion operation on the feature maps of different levels; performing further feature extraction on the fused feature data to obtain output layer data; analyzing the output layer data; and converting the target frame information into RGB data for visual display. The target detection effect can be improved, and the problems of large model size, large equipment weight, high system power consumption, dependence on a network and the like in the existing field environment crack detection technology are solved.

The application provides a SAR target scattering center parameter fast extraction method, and mainly solves the problems of slow speed and low precision of a traditional SAR target scattering center parameter extraction method; firstly, the application models a scattering center parameter solving problem based on a sparse representation theory; then, a SAR target scattering center parameter extraction problem model based on a semi-quadratic splitting method is constructed, a deep unfolding network is constructed according to an optimization process based on the semi-quadratic splitting method, and the SAR target scattering center parameter is extracted by using the deep unfolding network; compared with a traditional method, the method constructs a deep unfolding network based on the semi-quadratic splitting method and a scattering center model, and realizes efficient and interpretable SAR target scattering center parameter extraction.

The present invention provides an operator elimination method, system, electronic device and storage medium for a neural network, and relates to the technical field of deep learning. It includes: determining a target subgraph structure with operator elimination requirements in a neural network model; determining an operator in an eliminateable state in the target subgraph structure as a target elimination operator; traversing all subgraph structures until all target subgraph structures with operator elimination requirements and corresponding target elimination operators are determined; calculating the buffer memory sum of the input and output of the target elimination operator; determining the target cache area of ​​the target elimination operator according to the buffer memory sum, so as to allocate the target elimination operator to the target cache area, and configuring the output of all target elimination operators with the same cache address; determining the jump offset of the target elimination operator; compiling the target elimination operator based on the jump offset to complete operator elimination. The present invention can realize operator elimination and reduce the reasoning time of the neural network.

The invention provides a building worker action classification method of a hierarchical integrated network structure, which comprises the following steps: acquiring action data of a plurality of actions executed by a building worker wearing an IMU (Inertial Measurement Unit), preprocessing and establishing an action data set; selecting different IMU numbers and installation positions, sliding time window sizes and feature extraction data combinations to obtain a plurality of IMU configuration schemes; establishing a directed label tree based on the specificity among multiple actions, constructing a layered integrated network based on the directed label tree, and training the layered integrated network by utilizing action data in the action data set to obtain a trained action classification model; and inputting the action data of the complete action cycle of each IMU configuration scheme into the trained action classification model, evaluating the performance of each IMU configuration scheme according to the four accuracy performance indexes and the two real-time performance indexes, and selecting the IMU configuration scheme meeting a preset condition as an optional scheme for building worker action classification. The method has the advantages of being large in number of classified actions, high in precision, high in real-time performance and capable of providing various IMU configuration schemes, and it can be ensured that the wearable exoskeleton equipment accurately assists building workers in completing various operation tasks in real time.

This invention discloses a 3D multi-object tracking method based on spatiotemporal adaptive attention. The method includes: acquiring 2D multi-source data and 3D multi-source data of a target scene; extracting image features from the 2D multi-source data to obtain image features, and extracting point cloud features from the 3D multi-source data to obtain point cloud features; fusing the image features and the point cloud features to obtain fused features; determining the detection state of each object in the target scene based on the fused features; extracting cooperative features between the detected object and the predicted object based on the detection state; and updating the object detection and tracking state in the target scene in real time based on the cooperative features. This invention can effectively achieve better tracking performance and reduce inference time, and can be widely applied in the field of computer technology.

The invention discloses a scheduling method and device, processing equipment, a storage medium and a program product. The method comprises the following steps: receiving first input data of a user; performing topic tracking on the first input data to obtain a topic tracking result; determining a model scheduling demand according to the topic tracking result, and sending the model scheduling demand and encrypted target input data to a cloud side; wherein the model scheduling demand is used for the cloud side to schedule the encrypted target input data to a compression model or a full-amount model for reasoning.

The invention relates to memory allocation, in particular to an operator memory allocation method for complex structure deep learning model reasoning, and the method comprises the steps: counting the types of operators in a calculation graph; determining the memory allocation priority of each operator; determining the life cycle of each operator; selecting an operator which is not subjected to memory allocation and has the highest memory allocation priority, and performing memory allocation based on a storage structure of a deep learning acceleration kernel NNA and the life cycle of each operator; s4 is repeated until memory allocation of all operators is completed; according to the technical scheme provided by the invention, the defect that the reasoning time is difficult to effectively shorten when the method is applied to a deep learning model with a complex structure in the prior art can be effectively overcome.

This invention discloses an interpretable low-light image enhancement method. It constructs an unpaired learning framework, EDC-Net, with enhancement and degradation branches forming a closed loop. The enhancement branch adaptively enhances the brightness and contrast of the low-light image through pixel-level affine mapping, while the degradation branch predicts physically meaningful degradation parameters such as color gain and exposure scaling. The enhanced image is then reconstructed into a low-light image through a chain of explicit degradation operators, forming a cyclic consistency constraint. Training employs a three-stage progressive optimization strategy: degradation warm-up, master adversarial training, and fine-tuning. During the inference phase, only the lightweight enhancement branch is run. This invention achieves physical interpretability and controllability of the enhancement process, avoids overexposure and artifacts, improves training convergence stability, and significantly increases inference efficiency. It is suitable for real-time image processing scenarios such as nighttime surveillance and autonomous driving.

The invention provides an event camera real-time deblurring method based on dynamic edge guidance. The event camera real-time deblurring method comprises the following steps: acquiring a to-be-recovered blurred image and event stream data in the same field of view; extracting an image modal feature and an event modal feature by using a double-flow visual encoder; constructing a dynamic edge guiding module, predicting pixel-level multi-level fuzzy probability distribution, and generating a dynamic space edge priori graph; removing redundant background events by using an edge prior graph, and screening out a key event Token rich in motion information; and constructing a cross-modal fusion module based on bidirectional RWKV, carrying out long-distance space-time modeling on the screened event Token and image features with linear calculation complexity, and finally decoding and outputting a clear image. According to the method, the sensor noise is effectively suppressed and the calculation redundancy is reduced under the condition of not depending on the traditional hard gating through a'guide-selection 'normal form, and the image deblurring which is high in signal-to-noise ratio and high in structural similarity and meets the real-time requirement is realized.

The invention discloses an electric meter detection method and device based on SVTRLCNetV3. The method comprises the following steps: setting a plurality of robot inspection points; constructing and training a state value identification model and a numerical value identification model of the intelligent polling ammeter; making a template drawing and an algorithm configuration file of each robot inspection point; training a bidirectional LSTM sequence result analyzer; the inspection robot records an electricity meter video; according to the ammeter video, the algorithm configuration file corresponding to the robot inspection point and the template drawing closest to the current time period, executing an algorithm detection program to obtain an ammeter reading sequence; and a bidirectional LSTM sequence result analyzer is used to carry out reasoning on the ammeter reading sequence, and an ammeter value corresponding to each state value is output. According to the method, effective detection can be realized for the intelligent polling ammeter with multiple states continuously and circularly switched, an interference identification result is effectively eliminated, and the accuracy of the identification result is finally ensured.

The invention provides an autonomous navigation laser radar-camera fast fusion target detection method and device and a storage medium, and belongs to the field of multi-modal target detection, and the method comprises the steps: obtaining a laser radar point cloud and a camera image; obtaining point cloud BEV features according to the laser radar point cloud; performing feature extraction on the camera image to obtain a first image feature, and converting the first image feature into an image BEV feature by using a dynamic view angle converter; fusing the point cloud BEV features and the image BEV features to obtain fused BEV features; and a final target detection result is obtained. According to the invention, the operand of BEVPooling is reduced through the dynamic view angle converter, and the time and space complexity of the network structure is reduced, so that the reasoning time of the laser radar-camera fusion perception neural network is reduced, and the real-time performance of target detection is improved.

The invention relates to an operator generation method and device, equipment, a storage medium and a program product. The method comprises the following steps: constructing a deep convolutional neural network; determining a configuration item of the deep convolutional neural network and a plurality of kernel instances corresponding to the configuration item; running the deep convolutional neural network based on the kernel instance to obtain a speed measurement result corresponding to the kernel instance; based on the speed measurement result, selecting an optimal kernel instance from the multiple kernel instances; and generating a target operator based on the optimal kernel instance and the configuration item. According to the embodiment of the invention, the optimal kernel instance corresponding to each configuration item is determined by adopting an offline search mode for the multiple kernel instances corresponding to the common configuration items, and the corresponding optimal kernel instance can be directly called in a deep convolutional neural network reasoning process without online search, so that the reasoning time of a network model is shortened.

The application provides a real-time target detection method in an ultrahigh-resolution image scene, comprising: processing an original image to generate a high-resolution image and a low-resolution image; processing the low-resolution image to obtain a reserved area; cutting the high-resolution image according to the reserved area to generate a plurality of regional subgraphs; and detecting the regional subgraphs to obtain a detection result. The application can quickly detect the area requiring further fine-grained detection from an ultrahigh-resolution image through resolution processing, multi-scale density regression, area determination and area generation processing. In addition, due to the scale normalization, target detection and result merging processing, the final ultrahigh-resolution detection result can be generated. In the target detection stage, the plurality of regional subgraphs can be processed in parallel, which can greatly improve the inference time. Since part of the object-free area can be discarded in the density estimation stage, the required storage resources and calculation time can be saved.

The invention relates to a rapid pesticide residue screening method based on an optimized electronic nose array. The rapid pesticide residue screening method comprises the following steps: selecting a plurality of gas sensors made of different sensitive materials to form a basic sensor group, and performing array arrangement optimization; grading the to-be-detected agricultural product sample; synchronously collecting resistance change signals and response peak value occurrence time of the sensor in a specified time period; performing baseline correction on the dynamic response data, and then removing environmental interference noise; screening out specified key feature parameters; dividing the screened key feature parameters into a training set, a verification set and a final feature set according to a specified proportion, and performing training optimization verification on a fusion model of the support vector machine and the random forest; inputting the final feature set into the optimized screening model, and outputting a positive / negative judgment result and confidence; and respective processing is carried out according to the division of the confidence. Through sensor array optimization, pretreatment, multi-dimensional data preprocessing and fusion model construction, rapid and accurate screening of pesticide residues is realized.

The invention discloses a microscopic three-dimensional morphology reconstruction method of a high-speed deep network. The microscopic three-dimensional morphology reconstruction method comprises the following steps: acquiring N images of different focal planes of a microscopic scene at the same interval L; segmenting the N images according to the height and width of H * W to generate a tensor with the channel of N and the height and width of H * W; inputting the tensor into an encoder to obtain an extracted feature tensor; inputting the extracted feature tensor into a decoder to obtain a fused feature tensor; multiplying the fusion feature tensor by an interval L, and filtering to obtain a depth map; and re-splicing the depth maps according to the segmentation sequence. Through grouping convolution, the parameter quantity and the calculation quantity of a neural network model are effectively reduced, the inference time of the neural network is shortened, and real-time reconstruction of the microscopic three-dimensional morphology is realized; the information of the same size in the coding stage and the corresponding decoding information are added through the jump connection layer, gradient disappearance in neural network training is avoided, and the reconstruction precision of a microscopic scene is improved.