480 results about "Feature coding" patented technology

The invention relates to an infrared behavior identification method based on adaptive fusion of an artificial design feature and a depth learning feature. The method comprises: S1, improved dense track feature extraction is carried out on an original video by using an artificial design feature module; S2, feature coding is carried out on the extracted artificial design feature; S3, with a CNN feature module, optic flow information extraction is carried out on an original video image sequence by using a variation optic flow algorithm, thereby obtaining a corresponding optic flow image sequence; S4, CNN feature extraction is carried out on the optic flow sequence obtained at the S3 by using a convolutional neural network; and S5, a data set is divided into a training set and a testing set; and weight learning is carried out on the training set data by using a weight optimization network, weight fusion is carried out on probability outputs of a CNN feature classification network and an artificial design feature classification network by using the learned weight, an optimal weight is obtained based on a comparison identification result, and then the optimal weight is applied to testing set data classification. According to the method, a novel feature fusion way is provided; and reliability of behavior identification in an infrared video is improved. Therefore, the method has the great significance in a follow-up video analysis.

The invention discloses an attention mechanism relationship comparison network model method. An attention relationship comparison network model for small sample learning under a small amount of labeled sample data is constructed. Based on a relational network architecture, the model is divided into a feature coding part, a feature combination part and a relational coding part, the feature coding module is used for extracting image feature information, and the feature combination part is used for recombining the extracted query image feature information with the training image feature information of each group to form a new combined feature map. The relation encoding module performs nonlinear metric learning of the network; by introducing an attention mechanism and a spectrum normalizationmethod into an end-to-end deep convolutional neural network model, the model has higher classification accuracy under the condition of small sample learning, the stability of a final training result of the model is improved, and the image classification accuracy of an existing model in small sample learning is improved.

The invention provides a human body movement recognition method based on convolutional neural network feature coding and mainly aims to solve the problems of complicated calculation and low accuracy in the prior art. According to the implementation scheme, TV-L1 is utilized to obtain a video light steam graph; convolutional neural network coding, local feature accumulation coding, dimension-reducing whitening processing and VLAD vector processing are sequentially performed in a video space direction and a light stream movement direction, and space direction VLAD vectors and movement direction VLAD vectors are acquired; and information in the video space direction and information in the light steam movement direction are merged to obtain human body movement classification data, and then classification processing is performed. According to the method, convolutional features are subjected to local feature accumulation coding, so that the recognition rate is increased when complicated background data is processed, and the calculated amount is reduced; the features acquired by fusing video VLAD vectors and light stream VLAD vectors has higher robustness to environmental changes, and the method can be used for performing detection and recognition on human body movement in a monitoring video in areas such as a community, a shopping mall and a privacy occasion.

Feature extraction, coding and pooling, are important components on many contemporary object recognition paradigms. This method explores pooling techniques that encode the second-order statistics of local descriptors inside a region. To achieve this effect, it introduces multiplicative second-order analogues of average and max pooling that together with appropriate non-linearities that lead to exceptional performance on free-form region recognition, without any type of feature coding. Instead of coding, it was found that enriching local descriptors with additional image information leads to large performance gains, especially in conjunction with the proposed pooling methodology. Thus, second-order pooling over free-form regions produces results superior to those of the winning systems in the Pascal VOC 2011 semantic segmentation challenge, with models that are 20,000 times faster.

The invention discloses a context pyramid fusion network and an image segmentation method, and the context pyramid fusion network comprises a feature coding module which comprises a plurality of feature extraction layers which are connected step by step, and is used for obtaining a feature map of an original image; a plurality of global pyramid guiding modules, connected with the different featureextraction layers respectively and used for fusing the feature maps extracted by the feature extraction layers connected with the global pyramid guiding modules with the feature maps extracted by allthe higher feature extraction layers to obtain global context information and guiding and transmitting the global context information to the feature decoding module through jump connection; a scale sensing pyramid fusion module, connected with the highest feature extraction layer of the feature coding module and used for dynamically selecting a correct receptive field according to the feature maps of different scales and fusing multi-scale context information; and a feature decoding module, used for reconstructing a feature map according to the global context information and the multi-scale context information. The method is good in image segmentation performance, and is better in effectiveness and universality.

The invention discloses a spatial data conversion method and system. The method comprises: performing land feature coding, and marking CAD data by adopting GIS codes storing topographical elements; re-drawing graphic data by adopting features corresponding to graphs for the marked CAD data to normalize the CAD data; and converting the normalized CAD data into GIS data. A whole set of processing process is provided through an embodiment of the invention and comprises the functions of land feature batch coding, topologic check, specific graphic layer screening or exporting all data for the GIS and the like.

The utility model relates to a segmentation method of active web page, and is characterized in that the method first receives the content streams of web page and then builds a DOM tree, then makes the nodes of the DOM tree to feature codings, next compares the corresponding nodes of the DOM trees to build common blocks and customization blocks. The utility model can understand and identify the common parts (common blocks) sharing among multi-pages and the special parts of different variation rules (customization blocks) according to the dynamic characteristic and the structural characteristic of the web pages, and dynamically divide the web pages without human interference. The utility model provides a solution with a good expandability, lowers the labour costs for manual segmentation, and can be widely used in technology field of active web page.

The invention provides an end-to-end classification method of a large-scale news text based on Bi-GRU and a word vector. The end-to-end classification method comprises the following steps: S1. word Embedding word-level semantic feature representation is performed; S2. the attention weight Bi-GRU word level sentence feature coding model is constructed; S3. the Bi-GRU sentence level feature coding model based on the attention weight is established; S4. hierarchical Softmax is applied to realize end-to-end classification implementation. According to the method, the dimension of the vector can bereduced and the problem that the features are too sparse can be effectively prevented. The final output vector is optimized and the effectiveness of model feature coding is enhanced. The problem thatthe model is difficult to train because of the high dimension can be avoided and the additional semantic information can also be provided. The feature extraction model and various common classifiers can be flexibly combined so as to facilitate replacement and debugging of the classifiers. The computational complexity is reduced from | K | to log | K | in comparison with that of Softmax.

The invention discloses a road area detection method which comprises the following steps: S1, designing a road enhancement data enhancement method, and generating a road data enhancement function; S2,adopting a road data enhancement function, inputting a training sample, and outputting enhanced road data; S3, designing and training a feature coding network model, and outputting a coding feature map through the feature coding network model by using the enhanced road data; and S4, performing designing and training to generate a road segmentation decoding module and a road type classification decoding module, adopting the road segmentation decoding module and the road type classification decoding module, and inputting the coding feature map to output a road segmentation result and a road type classification result. The road area pixel-level segmentation and multi-type classification results are provided, which can be used for detecting passable areas of intelligent vehicles, and providesa basis for obstacle avoidance and path planning of the intelligent vehicles.

The invention provides a human-computer identification method based on a sliding trajectory, belongs to the technical field of computers and is used for solving the problem of low accuracy when a human-computer identification method simulates a sliding trajectory input by a user in the prior art. The method comprises the following steps that: determining the preset dimension feature coding of a historical sliding trajectory which is qualified, and the reference distribution feature of the feature coding of a historical sliding trajectory on the basis of at least one time window; determining the preset dimension feature coding of a real-time sliding trajectory to be verified, and the distribution feature to be verification of the feature coding of a historical sliding trajectory on the basis of at least one time window; and if the distribution feature to be verification on the basis of a certain time window is not matched with the reference distribution feature, determining that the real-time sliding trajectory to be verified is a machine simulation sliding trajectory. By use of the method disclosed by the embodiment of the invention, accuracy for identifying the machine simulationsliding trajectory can be effectively improved.

The invention belongs to the technical field of target detection, particularly relates to a three-dimensional target detection method, system and device based on RGB-D, and aims to solve the problem that efficiency and 3D target detection accuracy cannot be both considered in the prior art. The method comprises the steps of performing feature extraction on a 2D image set of a to-be-detected target, and reversely mapping the 2D image set to a 3D space; performing voxel division on the 3D point cloud data of the to-be-detected target, and performing feature coding through a 3D convolutional neural network in combination with the mapping voxel of the 2D image; aggregating 2D image texture features and 3D point cloud data geometrical features; obtaining a target feature cluster set through a Hough voting network; and obtaining a target bounding box as a three-dimensional target detection result through the target regression and classification network. According to the method, the 2D imagedata is reversely mapped to the 3D space, the 3D point cloud geometrical characteristics and the 2D image texture characteristics are fused, the 3D target detection accuracy is improved, the pre-selected area is generated through adoption of the Hough voting method, and the 3D target detection efficiency is guaranteed.

The invention relates to a user concentration degree identification method based on a hierarchical convolutional neural network. The method comprises the following steps that: obtaining the front side image of a face; according to the front side image of the face, utilizing two local binary pattern operators of an even pattern to calculate a feature coding graph corresponding to the front side image of the face; and according to the feature coding graph under two even patterns and the front side image of the face, adopting a GoogLeNet improved classifier to carry out classified processing to obtain the emotion of the user, and obtaining the user concentration degree according to the emotion. The invention also provides a user concentration degree identification system based on the hierarchical convolutional neural network. The user concentration degree result obtained by the invention is accurate and can be finely decomposed.

The invention discloses a method and a device for data processing based on federated learning, equipment and a medium. The method comprises the following steps executed by a first terminal: determining user feature data shared by a first terminal and a second terminal; performing feature coding processing on the user feature data to obtain to-be-processed feature data; obtaining a model predictionvalue obtained by processing based on the to-be-processed feature data; obtaining a loss value obtained by processing the training label data and the model prediction value by adopting a predefined loss function; sending the external loss value to the second terminal in an encryption mode if the loss value is the external loss value; and determining a target gradient based on the internal loss value and a current model parameter corresponding to the first to-be-trained model if the loss value is an internal loss value, and performing model optimization on the first to-be-trained model according to the target gradient to obtain a target prediction model. The data processing method based on federated learning can effectively improve the model training efficiency and accuracy.

The invention discloses a prediction method for protein post-translational modification methylation loci, and belongs to the field of bioinformatics. Protein methylation modification participates in cell functions and many life activities of cell processes, and recognition of protein methylation modification loci has very important significance in understanding of the life activities of cells. The prediction method combines with sequence information, evolutionary information and physical and chemical properties to conduct feature coding on a protein methylation sequence, an information gain optimization feature method is adopted and combines with a support vector machine to construct a prediction model, and it is shown through independent testing results that the prediction method has a good prediction property on the protein methylation loci; meanwhile, a network prediction platform is developed and used for conducting online prediction on the protein methylation loci.

The invention discloses a single-frame image super resolution reconstruction method based on sparse domain selection, mainly to solve the problem that the reconstruction result is poor caused when the existing reconstruction method carries out joint dictionary training. The method comprises steps: low-resolution and high-resolution image training sets are constructed according to an image set; low-resolution and high-resolution feature training sets are constructed according to the image training sets; sparse representation is carried out on the low-resolution feature training set; according to the high-resolution feature training set and a low-resolution feature coding coefficient, an iterative initial value for a high-resolution dictionary is solved; an optimization objective formula for sparse domain selection is constructed, and the high-resolution dictionary, a high-resolution feature coding coefficient and a mapping matrix are solved iteratively; and according to the inputted test image, the high-resolution dictionary, the high-resolution feature coding coefficient and the mapping matrix, a high-resolution image is reconstructed and outputted. The experimental simulation shows that the reconstruction result has higher subjective and objective quality evaluation, and can applied to medical imaging, high-definition video imaging, remote sensing monitoring, traffic and safety monitoring.

A feature coding unit extracts and encodes a feature of a video signal so as to generate a feature stream. A feature identifying unit checks a decoded feature obtained as a result of decoding the feature stream against a search key from a user for a match so that a video content requested by the user is retrieved.

The invention relates to a method for intelligently generating a power line route. The method comprises the steps of establishing rules: setting a feature coding rule, and establishing a basic line selection rule and a special line selection rule; establishing line selection regions, determining a target line selection region, and defining a working range; performing barrier analysis, obtaining geographic information data in the involved region according to the line selection regions, establishing a unit grid, judging whether the unit grid involved by geographic information is a barrier grid or not according to the line selection rule, and if yes, attaching attributes of the line selection rule to cells; and generating a line by a line search algorithm according to a cell set in the line selection regions. Compared with the prior art, the method for intelligently generating the power line route has the advantages that a line trend in an engineering region is comprehensively considered, scheme indexes are scientifically analyzed, and key index comparison and analysis are carried out, so that the route generation efficiency and accuracy can be effectively improved. In addition, the invention provides an apparatus for intelligently generating the power line route.

The invention discloses a three-dimensional point cloud target detection method. The method comprises the following steps: point cloud information of a three-dimensional scene is obtained through a depth sensor and an image sensor to serve as a training data set of a neural network; the point cloud of the target in the scene due to visual angle shielding and long-distance missing is complemented by utilizing a target point cloud model rendered by a computer; two three-dimensional target detection networks are constructed as a virtual training data set, one three-dimensional target detection network being used for inputting real data and the other three-dimensional target detection network being used for inputting virtual data, and the real three-dimensional scene point cloud data and the virtual three-dimensional scene point cloud data are respectively input into respective point cloud feature coding networks for feature extraction; the association perception process is simulated and applied to the deep neural network, and the incomplete point cloud information coding feature domain in the real scene is migrated to the virtual complete point cloud information coding feature domainthrough the transfer learning technology so that the neural network is enabled to actively associate the incomplete point cloud to the complete point cloud.

The invention discloses a palmprint recognizing method and device, relating to the field of biological feature recognition. The palmprint recognizing method comprises the steps of: sequentially intercepting a plurality of local areas on an area in which a palmprint image is interested and acquiring a contrast context vector of each local area; extracting stable areas from the plurality of local areas according to the contrast context vector of each local area; carrying out feature coding on pixels in the stable areas; and carrying out similarity judgment on palmprint features according to theresult of feature coding. According to the invention, prior to feature matching, the stable areas in the area in which a palmprint image is interested are extracted by using the contrast context vectors and the influence on a recognition result when unstable areas are matched is removed; and compared with the prior art, the palmprint recognizing method and device have the advantages that: the matching accuracy is higher and the influence caused by change of palmprint images is less.

An electronic cochlea Chinese fixed electrostimulation amplitude variation pattern exosomatic speech processing device comprises an audio amplification sampling module, a storage module, a digital signal processor and a signal transmission module, wherein the speech signal processing program of the device comprises a preprocessing unit, an endpoint detecting unit, a speech recognition unit and a feature coding unit; the feature coding unit has a fixed electrostimulation amplitude variation pattern library and a stimulation patter selecting and adjusting module; moreover, the feature coding unit selects a corresponding electrostimulation amplitude variation pattern from a fixed electrostimulation pattern library according to the recognition result of a speech section, and adjusts an electrode channel selection pattern, a stimulation speed variation pattern and stimulation time, thereby finally generating a holonomic electrostimulation parameter corresponding to each stimulation electrode. The electronic cochlea Chinese fixed electrostimulation amplitude variation pattern exosomatic speech processing device adopts a speech recognition technology which takes a Chinese standard syllable as a recognition unit, and carries out electrostimulation coding and adjusting of a recognition result by means of a fixed electrostimulation amplitude variation pattern, thereby restoring the Chinese speech recognition capacity of an electronic cochlea wear more effectively.

The invention discloses a video processing method and device, electronic equipment and a storage medium, and relates to the field of artificial intelligence, in particular to the fields of deep learning, model training, knowledge maps, video processing and the like. According to the specific implementation scheme, the method comprises the following steps: acquiring a plurality of first video frames, and performing fine-grained splitting on the plurality of first video frames to obtain a plurality of second video frames; according to multi-modal information related to the plurality of second video frames, performing feature coding on the plurality of second video frames to obtain feature fusion information used for representing multi-modal information fusion; and performing similarity matching on the plurality of second video frames according to the feature fusion information, and obtaining a target video according to a similarity matching result. By adopting the method and the device,the video splitting accuracy can be improved.

The invention discloses a continuous action online learning control method and system for an automatic driving vehicle. The continuous action online learning control method comprises the following steps: encoding a perceptual image It through a deep encoding network to obtain an encoding state feature st; respectively inputting encoding state features st into actuators-actuators, wherein the evaluator models all adopt an evaluator network and an actuator network of a cerebellar model neural network, an action at is output through the actuator network, and an actuator is updated through the evaluator network; parameters of an evaluator model. According to the invention, a synthetic deep neural network feature coding technology and an enhanced learning principle are adopted; the learning control problem of a continuous action space is solved under high-dimensional state input; on-line learning control of a continuous action space under large-scale continuous state input can be realized,the learning period is shortened while the learning effect is ensured, the learning process can be quickly converged to obtain a control strategy with a good performance effect, and the data utilization rate is good.