33results about How to "Strong expressive ability" patented technology

The invention provides a radar target recognition method based on micro-Doppler feature extraction and deep learning. The new method can be divided into three stages of step 1, a micro-Doppler featureextraction stage: using an MSTFT-WVD algorithm to obtain a micro-Doppler feature time-frequency graph of a target; step 2, a classifier model training phase: dividing the time-frequency graph of theknown category information obtained in the first stage into a training data set and a test data set, utilizing an original deep learning model for training, and continuously adjusting parameters to obtain the optimal deep learning model for radar target recognition; step 3, an identification phase of an unknown target: using the MSTFT-WVD algorithm to obtain a micro-Doppler feature time-frequencygraph of the unknown target, inputting the time-frequency graph to the trained deep learning model in the second stage to obtain the category information of the unknown target. The radar target recognition method can achieve higher recognition accuracy rate.

The invention provides a face recognition method. The method includes the steps of dividing a pretreated face sample image in fixed size into a plurality of superposed sub-regions in different sizes, extracting texture characteristics of the sub-regions, selecting the effective texture characteristics according to the preset rules, obtaining projection characteristic values of the effective texture characteristics, and carrying out face recognition according to the projection characteristic values of the sub-regions. The invention can improve the processing speed and the recognizing effect offace recognition.

The invention relates to an object six-degree-of-freedom pose estimation method based on color and depth information fusion. The object six-degree-of-freedom pose estimation method comprises the following steps of: acquiring a color image and a depth image of a target object, and carrying out instance segmentation on the color image; cutting a color image block containing a target object from thecolor image, and acquiring a target object point cloud from the depth image; extracting color features from the color image block, and combining the color features to the target object point cloud atthe pixel level; carrying out point cloud processing on the target object point cloud to obtain a plurality of point cloud local region features fusing the color information and the depth informationand a global feature, and combining the global feature into the point cloud local region features; and predicting the pose and confidence of one target object by means of each local feature, and taking the pose corresponding to the highest confidence as a final estimation result. Compared with the prior art, color information and depth information are combined, the object pose is predicted by combining the local features and the global features, and the object six-degree-of-freedom pose estimation method has the advantages of being high in robustness, high in accuracy rate and the like.

The invention discloses a pedestrian tracking method based on a low-altitude aerial photographing infrared video. Continuous and stable pedestrian tracking is realized through combination of a Lucas-Kanade optical flow method and local area secondary detection. The pedestrian tracking method comprises the steps that 1. an aerial photographing infrared pedestrian support vector machine (SVM) classifier is trained offline; 2. the initial position of a pedestrian target is determined; 3. the pedestrian target is preliminarily tracked by utilizing the LK optical flow method, and the position of the pedestrian target in the next frame is calculated; 4. a search area is set around the predicted position of the pedestrian target; and the infrared pedestrian is secondarily detected in the search area by utilizing the offline trained SVM classifier, and the position of the pedestrian target is updated; and 5. the center of the pedestrian target detected in the search area acts as input coordinates of the LK optical flow method of the next time, and the steps (3)-(5) are repeated. Continuous and stable tracking of the infrared pedestrian target can be realized by the pedestrian tracking method, and the problem of street lamp shielding can also be processed.

The invention provides a pest image automatic recognition method based on a high-reliability network. The method includes the following steps that preprocessing is carried out on multiple collected training images to obtain multiple training samples, and HOG feature extraction is conducted on the training samples to form joint image feature vectors of the training samples; the high-reliability network based on a restricted boltzmann machine is constructed, the joint image feature vectors of the training samples are input to the constructed high-reliability network, and training on the high-reliability network is completed; preprocessing is conducted on pest images to be tested to obtain test samples, and HOG feature extraction is carried out on the test samples to form joint image feature vectors of the test samples; the joint image feature vectors of the test samples are input to the high-reliability network after training is finished, and categories of past images to be tested are obtained through recognition. According to the method, the accuracy rate of pest recognition can be improved, and robustness of a pest recognition algorithm is enhanced.

The embodiment of the invention provides a medical knowledge graph fusion method and device based on multiple data sources. The medical knowledge graph fusion method comprises the steps of performingknowledge representation learning based on a first medical knowledge graph and a second medical knowledge graph, and obtaining each first initial vector and each second initial vector; mapping each first initial vector and each second initial vector into a reference vector space based on a pre-acquired reference vector set, and acquiring each first mapping vector and each second mapping vector; and according to each first mapping vector and each second mapping vector, fusing knowledge in the first medical knowledge graph and the second medical knowledge graph to obtain a fused medical knowledge graph. According to the medical knowledge graph fusion method and device based on the multiple data sources provided by the embodiment of the invention, the fused medical knowledge graph with stronger internal logicality and expression ability can be obtained by utilizing the logicality in the knowledge graph based on a knowledge representation learning method.

The invention discloses a motion synthesizing and editing method based on motion capture data in computer bone animation. The method comprises the following steps: simply preprocessing the given motion data set, and marking required prior information for constructing and generating a model; defining multi-factor Gaussian random process according to the marked information to model the motion data set; constructing an objective function based on MAP (maximum posterior estimation) according to the constructed model, and using SCG (scaled conjugate gradient) optimization algorithm to solve the unknown function to obtain the generative model; and at last, editing hidden variable factors, and utilizing the obtained generative model to realize various motion synthesizing and motion editing methods, for example: style transfer, style interpolation and motion redirection. The method has the characteristics as follows: the multi-factor Gaussian process is utilized to model a group of less motion data to obtain the generative model, and the generative model is utilized to realize a series of interactive and visual motion editing methods.

The embodiment of the invention provides a drowning behavior identification method, a monitoring camera and a monitoring system, and the method comprises the following steps: collecting a plurality offrames of images in a monitoring water area; inputting a single-frame image in the multiple frames of images into a preset deep neural network model, obtaining the drowning confidence of each human body target in the single-frame image, wherein the preset deep neural network model is a deep neural network model obtained by training in advance according to a sample image containing the key point position information of the human body target collected from the upper part of the water area when the human body target is drowned; and for each human body target, identifying whether the human body target has a drowning behavior or not according to the drowning confidence of the human body target. Through the scheme, the drowning behavior recognition precision can be improved.

The invention relates to an image retrieval method based on multi-scale NetVLAD and depth Hash. According to the method, a multi-scale convolutional neural network based local aggregation descriptor vector method is adopted to carry out feature extraction on pictures in image retrieval. The original output features which only adopt the last convolutional layer are optimized into the features whichare fused by adopting the output features of the plurality of convolutional layers. The fused features not only comprise high-layer semantic features, but also comprise low-layer picture detail information. A Hash layer is added behind the NetVLAD for feature coding, the features become simpler, the feature dimension and complexity are reduced through a Hash coding layer, and the subsequent storage overhead and calculation complexity are remarkably reduced. The speed of image retrieval is increased, and similar images can be quickly and accurately retrieved in a large-scale data set.

The invention discloses an ontology data storage method and relates to the technical field of ontology data storage.The method is mainly used for effectively and reasonably storing concepts (or examples) and relations (or attributes) between the concepts.A data storage mode of a Nosql database is adopted to a database so that semantic information of data can be conveniently stored, processing of data between ontologies and calculation and reasoning between ontology data are facilitated, and the bottleneck problem about storing and obtaining the ontology data through RDF, OWL and a relation database is solved.The method meets the requirement for separated storage of symbolic languages and object IDs, effectively solves the problem that object terms and objects are always mixed, can be applicable to storage of ontology data in any field, can be compatible with existing ontologies in other formats, and can be used for storage of knowledge maps.

The invention provides an operation action identification method and device. The method comprises the steps that a to-be-identified video clip is acquired, wherein the to-be-identified video clip contains an action type; and the action type in the to-be-identified video clip is identified according to the to-be-identified video clip and a pre-established action identification logo model. Through the operation action identification method and device, information can be extracted layer by layer from pixel-level original data to abstract semantic concepts, extracted features have more efficient expression capability compared with manually designed features, and an operation action can be identified quickly and accurately.

The invention provides a multi-sample facial expression recognition method based on low-rank tensor decomposition. The multi-sample facial expression recognition method comprises an image preprocessing step, a feature extraction step, a tensor modeling step, a low-rank learning step, a tensor decomposition step and a feature classification step. According to the multi-sample facial expression recognition method, tensor representation feature space is utilized to reserve nonlinear features of an image; and face sub-space area features of different individuals are learned through a low-rank tensor decomposition technology, and face information under different dimensions is obtained, and tensors under all sub-spaces are decomposed, clustered and reconstructed to obtain effective expression features, and the expression information expression capacity is higher, so that the face expression recognition rate is increased.

The invention relates to the field of unconstrained facial expression recognition, in particular to a lightweight unconstrained facial expression recognition method and system embedded with high-order information, and the method comprises the steps: carrying out the preprocessing and image enhancement of input data, inputting the data into a lightweight feature extraction network, and extracting a deep feature map of a facial expression image; inputting the deep feature map into the input of a second-order effective channel attention module, counting the second-order information of the deep expression features and capturing the interdependence relationship between the cross-channel features; using cross entropy loss and center loss to jointly optimize the network model; inputting a to-be-detected facial expression image into the trained network model, and outputting a final predicted expression category by the classifier according to facial expression features; the network model provided by the invention has less parameter quantity, lower video memory requirement and calculation amount, does not use additional data pre-training models, is higher in precision, and is higher in applicability of related products.

The invention relates to a rapid production method of yarn-dyed fabric based on a cloud platform. The method includes: uploading an image of the produced yarn-dyed fabric A determined by category andproduction information to a cloud platform, selecting an image of A matched with the yarn-dyed fabric B to be produced and producing B with reference to the production information of A, the matching is that the image of B is inputted to the identification network corresponding to the class B to obtain the code of B after the class B is determined, selecting an image of A corresponding to the codeexactly consistent with the code, The construction process of the identification network is as follows: with the image and encoding of A in the same category as input and output, Firstly, all parameters of the VGG network are adjusted continuously, and the relatively optimal generation parameters are selected to substitute into the VGG network to obtain the pre-identification network. Secondly, the parameters of the whole connection layer network of the pre-identification network are adjusted continuously, and the relatively optimal generation parameters of the whole connection layer network are selected to substitute into the pre-identification network to obtain the identification network. The invention realizes high-precision and high-efficiency on-line retrieval of yarn-dyed fabric, shortens the production cycle and saves the cost.

The invention discloses a pedestrian re-identification method, a pedestrian re-identification system, a pedestrian re-identification device and a storage medium. The method comprises the steps of firstly, training an auto-encoder model by utilizing a pedestrian image database, optimizing a weight matrix, and obtaining an optimal model; calculating the feature score of each image by using the weight matrix of the hidden layer of the trained model, and sorting the features from high to low according to the scores to obtain high-score features; and 3, fusing the first three features of the scorein parallel to obtain the fusion feature with strong expression capability and low dimension; and finally, querying the similarity between the image and one target sequence in a database, and sortingaccording to the similarity from high to low to obtain an optimal matching result. Through the combination of the self-coding unsupervised feature selection and the parallel feature fusion, the unsupervised pedestrian re-identification is realized, the manpower and material resource input can be reduced, and the target searching task efficiency is improved.

The invention discloses a high-precision farmland vegetation information extraction method, and relates to the technical field of agricultural remote sensing, and the technical scheme is characterized in that the method comprises the following steps: S1, collecting an original farmland image; S2, performing flat field correction processing on the farmland original image; S3, importing the image data after the flat field correction processing into Pix4Dmapper or ENVI software, and carrying out the image splicing and cutting; S4, labeling the image data, and formulating a neural network data set through a preprocessing mode of data enhancement and image cutting; S5, constructing and training a Unet neural network model; S6, saving the model, inputting a farmland image to be identified into the saved model, and obtaining an extracted farmland vegetation texture and spatial distribution information result. According to the method, the texture and spatial distribution information of the crops can be quickly and accurately obtained from the remote sensing image of the farmland, the problems of complex manual screening features and low recognition precision existing in satellite remote sensing interpretation of the farmland image are solved, and a reference method is provided for remote sensing interpretation of farmland crop information by an unmanned aerial vehicle.

The invention discloses an industrial part key point detection method based on deep learning in the field of industrial vision. The method specifically comprises the following steps that S1, a key point detection deep neural network is trained, specifically, the key point detection deep neural network composed of three sub-networks is constructed, each sub-network comprises a plurality of 3 * 3 convolution kernels with the step length ranging from 1 to 2, and feature fusion is conducted between the sub-networks through the multi-size feature map fusion technology; S2, outputting a to-be-detected image to the training key point detection deep neural network, key points are detected by using the key point detection deep neural network obtained by training; the method is applied to a convolutional neural network structure for detecting the key points of the bearing workpart, and the network structure comprises a novel feature map fusion technology and a target loss function, so that the key point detection accuracy can be effectively improved.

The invention relates to a dynamic frame interpolation technology-based curved surface process time and space simulation method. The method comprises the following steps of: starting simulating; constructing a simulation initial curved surface; initializing parameters of the curved surface simulation; constructing dynamic frame interpolation; and performing three-dimensional display on the curvedsurface constructed by the frame interpolation. The step of constructing the dynamic frame interpolation comprises the following steps of: calculating time values, calculating Z values of all points in a frame interpolation way, updating the Z values of all the points, and reconstructing again. According to the method, a curved surface space model is constructed in each frame interpolation time and is displayed by continuously calculating the time values until the last curved surface model is displayed. By the dynamic frame interpolation technology, the problem that the simulation process hops and is not fluent when a sample observation value is relatively limited, the conventional expression way of observation information is broken through, the observation information is expressed by pictures with curved surface time and space change, the change process is vividly and directly simulated, and the method is especially suitable for the simulation of the change process of the curved surfaces of groundwater simulation, ground subsidence and the like.

The embodiment of the invention discloses a gender and age identification method and device, a storage medium and a server, and belongs to the technical field of video analysis. The method comprises steps that according to N kinds of cutting rules, face images of a to-be-identified pedestrian are respectively cut to acquire N local face images, whereinN is a positive integer greater than 1; for each local face image, the local face image is input into a recognition model corresponding to the cutting rule of the local face image in an optimal model combination, the optimal model combination comprises N recognition models with different model structures, and each recognition model corresponds to one cutting rule after being trained; and gender information and age information of the pedestrian are determined according to an output result of the optimal model combination. According to the embodiment of the invention, the accuracy of gender identification and age identification can be improved, and the storage space and program running time are reduced.

The invention relates to a branch CNN-based local climate region classification structure through using SAR and multispectral remote sensing data. The branch CNN-based local climate region classification structure comprises an SAR data preprocessing module and a multispectral data preprocessing module. The SAR data preprocessing module and the multispectral data preprocessing module are respectively connected with respective corresponding Inception-based branch CNN modules; wherein the two Inception-based branch CNN modules are both connected with an enhanced dense convolutional network, and the enhanced dense convolutional network is sequentially connected with a full connection layer and a softmax layer which are used for outputting local climate region categories. The invention furtherprovides a local climate region classification method using SAR and multispectral remote sensing data based on the branch CNN. According to the method, the integration of SAR and multispectral data can be realized, and the branch CNN can be used for respectively extracting the characteristics of the SAR and multispectral channel data, so that the multi-source remote sensing data is fully utilized,and the problem of local climate region classification based on the remote sensing data is better solved.

The invention belongs to the technical field of gas exploitation, and particularly relates to a coal and gas outburst strength prediction method based on deep learning, which comprises the following steps: step 1, data preparation: selecting prediction indexes of coal and gas outburst, defining training data, and standardizing the data; 2, feature extraction: defining network or model compositionaccording to a data set, mapping input to a target, and extracting geological index features; 3, configuring a learning process, selecting a loss function, an optimizer and an index to be monitored, and setting the number of iterations; 4, training a model, inputting a fit method of a sample calling model to iterate on training data, and training and optimizing the model; and 5, verifying the model, predicting coal and gas outburst samples on the verification set, comparing the predicted coal and gas outburst samples with actual results, and determining the prediction precision of the model, so that the structure is reasonable, the expression ability is stronger, the mapping effect is better, and the outburst strength prediction accuracy can be further improved.

The invention discloses a data category attribute representation method and an access control method. The method comprises the following steps of: 1) classifying data resources of each layer of a target information system according to a set classification theme x to obtain a data classification tree Tx or a classification forest consisting of a plurality of trees; 2) for a data resource xi in thesystem, querying a corresponding data classification tree Tx, obtaining a path from a Tx root node to a node Txi corresponding to the xi, and representing an attribute value of the xi corresponding tothe classification theme x by utilizing the path information; 3) setting an attribute value of each main body in each layer of the target information system according to an access control requirement, and determining an attribute range of each main body capable of processing data resources; and 4) when the subject accesses the object of the corresponding layer, determining whether the attribute value of the subject is matched with the attribute value of the object or not, and if the attribute value of the subject is matched with the attribute value of the object and meets the set access condition, allowing the subject to access the object, the object being the data resource of each layer of the target information system.

The invention relates to a community discovery method based on nonlinear non-negative matrix factorization, which comprises the following steps of: S1, formalizing a social network to generate an undirected graph; s2, according to the undirected graph, introducing a graph convolutional network GCN and an NMF model to construct a unified loss function; s3, optimizing the unified loss function to obtain a local optimal solution; and S4, extracting a community discovery result according to the local optimal solution. The graph convolutional network GCN and the NMF model are introduced to construct the unified loss function, and joint training is performed to endow the NMF with a nonlinear feature learning ability, so that the performance of the NMF-based community discovery method is improved.

The invention discloses a license plate classification and recognition method based on a Gabor feature auto-encoder. The method comprises the following steps: step 1, preprocessing a license plate image; step 2, performing pyramid zooming processing on the preprocessed license plate image; step 3, carrying out Gabor transformation on the license plate image subjected to zooming processing in the step 2; step 4, carrying out local sampling on the license plate image subjected to scaling processing in the step 2 to obtain a license plate image block, and carrying out local sampling on the license plate image subjected to Gabor transformation in the step 3 to obtain a local Gabor feature block; step 5, inputting the license plate image block and the local Gabor feature block obtained in the step 4 into an auto-encoder network with Gabor features, and solving parameters in the auto-encoder network structure with Gabor features through a back propagation algorithm; step 6, extracting license plate features; and step 7, performing license plate classification and recognition for the license plate features obtained in the step 6. According to the method, the over-fitting problem can be solved to a certain extent, and the license plate recognition accuracy is improved.

The embodiment of the application discloses a gender and age identification method, device, storage medium and server, belonging to the technical field of video analysis. The method includes: according to N kinds of clipping rules, respectively clipping the face images of pedestrians to be recognized to obtain N partial face images, wherein N is a positive integer greater than 1; for each partial face image, The partial human face image is input into a recognition model corresponding to the cropping rule of the partial human face image in the optimal model combination, the optimal model combination includes N recognition models with different model structures, and each recognition model After the model is trained, it corresponds to a clipping rule; the gender information and age information of the pedestrian are determined according to the output result of the optimal model combination. The embodiments of the present application can improve the accuracy of gender identification and age identification, and reduce storage space and program running time.

The invention discloses a link prediction method of a graph attention network based on node similarity, which comprises the following steps: firstly, constructing an undirected heterogeneous network model, preprocessing node data, and calculating a similarity matrix of the network; secondly, constructing a graph attention network model based on a node similarity index, and solving embedded representations of all nodes in the network; thirdly, carrying out negative sampling on information of existing edges, dividing all true edges and false edges into a training set and a test set, and using the training set and the test set in model training; and finally, inputting the embedded representations of the nodes learned through a graph attention method based on node similarity into a machine learning classifier in a series connection manner, and performing link prediction. According to the method, more related nodes are linked together through the node similarity index, so that the embedded representation learned by each node more fully represents the structural information of the network, and the embedded representation of the node obtains a better result in a downstream link prediction task.

The invention discloses a pedestrian detection method and system based on dynamic grouping convolution, belongs to the technical field of neural networks, and aims to solve the technical problem of how to realize pedestrian detection by utilizing grouping convolution and effectively improve the pedestrian detection efficiency. A target detection network architecture is constructed by using packet convolution, and initial training is completed while an original neural network structure is reserved; calculating scores of the input channel and the output channel by using a significance score generator, and sorting according to a sequence from high to low; setting a threshold value, and screening an input channel corresponding to the output channel; replacing the last 1 * 1 point-by-point convolutional layer in each inverted residual block by using a channel selector; training the target detection network architecture again by using the pedestrian data to obtain a reasoning model; and deploying the inference model at a side end, and identifying and judging a target acquired in real time.

The invention relates to a multi-dimensional expansion prediction method and device for non-stationary time series data, and the method comprises the following steps: collecting original meter readingdata, processing the original meter reading data, and expanding data dimensions to construct a meter reading data set; constructing RNN sequence expression models with different sequence dependent lengths, carrying out model training by adopting total data, and determining the optimal dependent length according to model performance expression; establishing an RNN sequence model based on the optimal sequence dependence length, and performing model training based on the meter reading data set to obtain optimized RNN sequence model parameters; predicting the latest remote transmission data of each meter by using the optimized RNN sequence model parameters; and comparing the predicted remote transmission data value with the actual remote transmission data value, obtaining a correction value according to a strategy, and recording the correction value into a database. According to the method, the non-stationary time sequence can be modeled, and meanwhile, the information dimension of the difference time step length can be fused, so that the time sequence model has stronger expression capability.

The invention provides a user risk assessment method and device, electronic equipment and a storage medium. In the user risk assessment method, user data of a to-be-assessed user is firstly obtained, and the user data at least comprises user attributes, relation information and transaction information. And inputting the user attributes and the relationship information into a knowledge extraction unit to obtain user knowledge features. And inputting the transaction information into a first linear mapping unit to obtain user transaction characteristics. Inputting the user knowledge features and the user transaction features into a dynamic fusion unit to obtain fusion features; and finally, inputting the fusion feature into a second linear mapping unit to obtain the risk level of the user to be evaluated. The problems that in the prior art, the risk assessment effect of the user is poor, the risk of the user cannot be accurately assessed, and the risk of financial services of financial institutions such as banks is increased are solved.