32results about How to "Improve expression accuracy" patented technology

The invention discloses a method for real-time monitoring of the face image quality of a customer collection terminal, comprising the following steps: Step 1, face image collection; Automatically detect and segment the exact position of the face in the face image; step 3 comprehensively evaluates the face quality, and performs intelligent prompts and real-time intelligent monitoring of the image quality through the face image quality detection system for the acquisition process of the client's face image. The invention can effectively improve the accuracy of subsequent image comparison and reduce the false alarm rate (False Alarm Rate), thereby ensuring the reliability of identification information and making the intelligent monitoring system based on face recognition truly practical.

The invention discloses a named entity recognition model training method, a named entity recognition method and a named entity recognition device. The training method comprises the following steps: preprocessing a corpus sample to obtain a character sequence sample, and labeling a named entity label on the character sequence sample to obtain a training character sequence; pre-training the trainingcharacter sequence based on a first bidirectional language model and a first self-attention mechanism model to obtain a character feature vector and a character weight vector, and fusing the character feature vector and the character weight vector to obtain a second context vector; pre-training the training character sequence based on a second bidirectional language model and a second self-attention mechanism model to obtain a word feature vector and a word weight vector, and fusing the word feature vector and the word weight vector to obtain a second context vector; and training the bidirectional neural network and the conditional random field which are connected in sequence by using the first context vector and the second context vector to obtain a named entity recognition model. According to the method, the training effect of the named entity recognition model is effectively improved, and the named entity recognition accuracy is improved.

The invention discloses a content complexity adaptive video satellite compressed image super resolution method comprising the following steps: dividing an observation image into zones of unequal content complexity from the texture thickness and structure simple and complication angles; gathering image samples with the consistent characteristics so as to form training image sets; forming image sets of different attributes, and using each image set to train a depth learning network model; using the model matching with image different zone statistics characteristics to carry out super resolution reconstructions of the corresponding zones. The method considers content complexity differences of satellite image different natural object types, thus effectively improving the video satellite image super resolution reconstruction precision.

The invention discloses a remote-sensing artificial ground object identifying method based on a semantic tree model of an object. The remote-sensing artificial ground object identifying method comprises the steps of: establishing a remote-sensing ground object representative image set; splitting images in the remote-sensing ground object representative image set by adopting a multi-scale method, and obtaining an object tree of each image; modeling for each node of each object tree by adopting an LDA (linear discriminant analysis) method, and computing implied semantic features contained in the tree node objects; obtaining the object tree sets of all the images in the representative set to learn each pair of object trees in a matching way, and extracting the common maximum sub-trees from the object trees; combining all the common maximum sub-trees together by adopting a step-by-step adding method, and forming an object semantic tree of the category of the described ground object; and identifying the artificial ground object according to the object semantic tree and obtaining the area in which the ground object is positioned. The remote-sensing artificial ground object identifying method disclosed by the invention can be used for mostly effectively processing the artificial ground objects in the condition of high-resolution remote-sensing images; the identification result is accurate, the robustness is good, the applicability is high, and manual work is reduced.

The invention provides a method for backward deducing a constitutive model of a welding joint based on a nanoindentation test, comprising the following steps: step 1, constructing a constitutive modelof the welding joint; 2, carry out dimensional analysis to obtain a dimensionless function between stress and strain; 3, selecting that material satisfying the constitutive relation of the step 1, and carry out finite element simulation on the process of pressing the indenter into the material; 4, determining the selected material and the corresponding stress and strain, substituting the selectedmaterial into a dimensionless function, and determining the function value corresponding to each material; 5, fitting the obtained mechanical property parameter and the function value to obtain the expression of the dimensionless function; 6, perform nano indentation test on that weld joint to obtain the mechanical property parameters of the weld joint; Step 7: The constitutive model of the welded joint is deduced backwards by using the mechanical property parameters. The invention provides a reverse inference method for welding joint constitutive model based on nano-indentation test, which has the advantages of wide applicability, low cost and high accuracy.

An expression estimation device configured to estimate an expression of a person from an image has an expression determination unit that analyzes the image of the person to generate first expression information indicating an expression of the person, a motion determination unit that analyzes the image of the person to determine a type of motion of the person, and an expression estimation unit that generates second expression information indicating the expression of the person according to the first expression information generated by the expression determination unit and the type of motion determined by the motion determination unit.

The invention provides a satellite video super-resolution reconstruction method and system based on a recurrent neural network. The method comprises the steps: obtaining a satellite video image, carrying out the frame-by-frame extraction, and forming an image sequence of the same region; carrying out multi-scale object segmentation on the extracted image sequence according to the shape and texturefeatures of the ground object to serve as geometric constraint conditions to ensure the integrity of the ground object, and obtaining an image sequence after multi-scale segmentation, denoted as theimage sequence as a high-resolution image sequence; simulating degradation processing in a satellite video image transmission process, and performing down-sampling, compression and noise addition on the image sequence after scale segmentation to obtain a low-resolution image sequence after degradation; constructing a deep learning network by taking the low-resolution image sequence as input and taking the current frame image as output based on a recurrent neural network; and optimizing the network, and outputting a super-resolution reconstructed video image by using the trained recurrent neural network. According to the invention, the integrity of ground objects is ensured, and the reconstruction precision of satellite video images is improved.

The invention provides a long text matching method and device, electronic equipment and a storage medium and relates to the technical field of artificial intelligence. The long text matching method comprises the steps of obtaining first long text data and second long text data; constructing a first co-occurrence matrix corresponding to the first long text data, and constructing a second co-occurrence matrix corresponding to the second long text data; constructing a similarity interaction matrix corresponding to the first long text data and the second long text data based on the first co-occurrence matrix and the second co-occurrence matrix; and performing conversion processing on the similarity interaction matrix to obtain a matching result corresponding to the first long text data and thesecond long text data so as to select target long text data matched with the first long text data from the second long text data according to the matching result. According to the method and the apparatus, the long text data with relatively high similarity can be matched for the input long text data, so that the accuracy of the matched long text data is improved.

The invention relates to a patient behavior multimodal analysis and prediction system based on statistical learning. The system comprises a multimodal acquisition module, a posture-based patient behavior recognition module, a physiological signal-based patient behavior recognition module, an emotion signal-based patient behavior recognition module, a voice signal-based patient behavior recognitionmodule and a multi-kernel learning-based fusion module. The multi-kernel learning-based fusion module comprises a multi-kernel classifier, and the training process of the multi-kernel learning-basedfusion module specifically comprises the steps that all kernel functions are trained and combined, then overall training is conducted, and the weight coefficient of each kernel function is obtained. Compared with the prior art, according to the invention, the applied multimodal data is closer to the real form of the patient information flow under the background of cloud computing and big data, hascomprehensiveness and complexity, reduces the loss of a large amount of information when facing complex nonlinear multi-modal information processing by utilizing the multi-feature multi-core learningmethod, and has better performance when processing data with larger modal span.

The invention relates to an electric vehicle charging load prediction method based on dynamic energy consumption and user psychology. The method comprises the following steps: setting the number of electric vehicles in a to-be-predicted area and an initial SOC; constructing an electric vehicle travel model and an actual power consumption model; according to the travel model, obtaining the optimal strategy of each electric vehicle through a Markov dynamic path decision model, and simulating the travel process of each electric vehicle according to the corresponding optimal strategy; in the simulation process, for each travel road node of each electric vehicle, the power consumption of the next travel process of the electric vehicle and the current SOC are calculated in advance according to the actual power consumption model and the initial SOC, and the charging demand of the electric vehicle at the current travel road node is determined. And accumulating the charging demands of all the electric vehicles at each travel road node, and obtaining the space-time distribution of the charging demands of the electric vehicles in the to-be-predicted area. Compared with the prior art, the method has the advantages of high accuracy, high reliability and the like.

The invention discloses a pre-training method and device for a semantic representation model, electronic equipment and a storage medium, and relates to the field of artificial intelligence. The specific implementation scheme is as follows: collecting a plurality of single-mode training data sets with different modes and a multi-mode training data set comprising a plurality of modes at the same time; respectively adopting the training data set of each single mode to pre-train the semantic representation model of the corresponding single mode; and based on the trained semantic representation model of each single mode, adopting a multi-mode training data set to train the fused multi-mode semantic representation model, wherein the multi-mode semantic representation model comprises a pluralityof single-mode semantic representation models and a fusion model. According to the method, the multi-modal semantic representation model can be trained in stages, and the expression accuracy of the multi-modal semantic representation model can be effectively improved.

The invention provides an information recommendation method and device, equipment and a storage medium. The method comprises the steps of receiving a recommendation request, wherein the recommendation request comprises an identifier of a user; determining a candidate object; according to the identifier of the user and a preset corresponding relationship, determining a user feature vector corresponding to the user, the corresponding relationship being used for representing a mapping relationship between the user and the user feature vector, and the user feature vector being obtained by training vectorized representation of the user in the user data; the vector is used for representing the interest degree of the user on the object and the influence degree of the interest of the social friends of the user on the interest of the user; according to a user feature vector and a feature vector of the candidate object, the recommendation degree of the candidate object to the user is determined, and the feature vector of the candidate object is obtained by training vectorized representation of the object in the user data and is used for representing a vector of a user group of the object; and sorting the candidate objects according to the recommendation degree to obtain a recommendation result, and returning the recommendation result to the terminal equipment of the user.

The invention relates to a face super-resolution processing method and system based on the closed link data reexpression. The method comprises the following steps of firstly, constructing a training library comprising a high-resolution human face image library and a low-resolution human face image library corresponding to the high-resolution human face image library; secondly, dividing the low-resolution face image to be processed and the image in the training library into the image blocks with overlapped parts by adopting the same blocking mode; pre-training the global correlation degree of each block position in the low-resolution space and the high-resolution space; then performing data re-expression in the low-resolution image space on an input low-resolution image to be processed; performing the data expression from the low resolution space to the high resolution space on the input image to be processed after the low resolution space is reexpressed; carrying out data re-expressionof the high-resolution space; and finally, splicing the high-resolution face image blocks to obtain a high-resolution face image. The method and the system are particularly suitable for recovering the face image in a low-quality monitoring video.

The invention relates to a touch type scene icon displaying method of a mobile terminal. A function icon is displayed by adopting the material object pattern mode similar to actual life, expression accuracy of the pattern function is greatly improved, user recognition speed and accuracy are improved, and the touch type scene icon displaying method is especially suitable for being used by the aged with poor sight and people knowing few words and poor in recognition.

The invention provides an incremental learning method for multi-frame image spectrum dictionary construction, and relates to the technical field of remote sensing image processing, and the method comprises the steps: obtaining a plurality of frames of spectrum images; performing dictionary learning by using the first frame of spectral image to obtain a spectral dictionary expressing the first frame of spectral image; dividing the spectrum dictionary expressing the first frame of spectrum image into a universal spectrum dictionary and a specific spectrum dictionary of the first frame of spectrum image through the activity ratio; solving a specific spectrum dictionary of the second frame of spectral image by using the specific information of the second frame of spectral image and an outlier removal strategy; using common information of the first frame of spectral image and the second frame of spectral image to update the universal spectral dictionary; and performing iterative processing on the multiple frames of spectral images to obtain an incremental dictionary composed of the specific spectral dictionary of the multiple frames of spectral images and the updated universal spectral dictionary. According to the invention, while the original spectral image dictionary expression performance is considered, the dictionary expression precision of the newly added spectral image is effectively improved.

The invention discloses a three-dimensional MRI (Magnetic Resonance Imaging) lesion image segmentation method and a three-dimensional MRI lesion image segmentation system. The three-dimensional MRI lesion image segmentation method comprises the following steps: S1, obtaining an important lesion feature image sequence from a multi-modal MRI image under various weighted S2, sequentially performing focus feature reconstruction on each group of focus important feature image sequences to obtain MRI focus images in corresponding weighted imaging modes, and sequentially calculating focus information degree of each MRI focus image; and S3, constructing a modal weight corresponding to a weighted imaging modal based on the lesion information degree, and performing modal feature reconstruction on all the MRI lesion images according to the modal weight to obtain a three-dimensional MRI lesion image. According to the three-dimensional MRI lesion feature representation method, all lesion features in various weighted imaging modes are subjected to generalization and personalized overall representation in the three-dimensional MRI lesion image, so that the expression accuracy of the lesion features of the target object is improved, and the auxiliary precision of a doctor on disease diagnosis is improved.

The invention relates to the field of engineering machinery, and discloses a construction site scene image acquisition system and method and a construction site object identification system and method. The construction site scene image acquisition system comprises: a camera shooting unit which is arranged on a cantilever crane of engineering machinery and is used for capturing a plurality of single images of a construction site area by adapting to different cantilever crane motion states; a cantilever crane detection unit which is used for detecting a plurality of rotation angles or a plurality of pieces of cantilever crane attitude information corresponding to different cantilever crane motion states, including the corresponding rotation angles when the cantilever crane moves; and an image processing unit which is used for carrying out angle correction on the corresponding single image according to the rotation angle and splicing a plurality of single images subjected to angle correction to form a construction site scene image representing the construction site area. According to the invention, the camera shooting unit is arranged on the arm support, so that image acquisition at a top view angle is easy to achieve, mutual shielding among objects on a construction site is avoided, and an image correction scheme based on a rotation angle is provided, so that high-quality images can be formed.

The invention provides a seamless fusion expression and correction method of global navigation satellite system (GNSS) global and regional ionospheric delay. Ionospheric delay above regions are expressed as a series of satellite-observation station pair elementary units, and new ionospheric delay quantities provided by newly added observation stations or satellites are used as independent units and directly added to the tail of a sequence so that seamless fusion expression of the ionospheric delay in any range is achieved; the inside of a single observation station-satellite pair unit is expressed as a satellite number plus a puncture point azimuth angle plus a puncture point elevating angle plus slant path ionospheric delay plus slant path ionospheric delay accuracy, and users select a plurality of the most matchable slant path ionospheric delay pairs interpolated by ionospheric delay correction quantity of the users according to the satellite number, the puncture point azimuth angle and the puncture point elevating angle after receiving sequences of slant path ionospheric delay pairs. The seamless fusion expression and correction method retains ionospheric delay expression and correction accuracy in an uttermost mode and achieves seamless fusion of global ionospheric expression and correction models.