75results about How to "Generalizable" patented technology

The invention relates to a video generation model based on text description information and a generative adversarial network. According to the model, videos with the text description information are used as training data; a self-help sampling method is adopted to extract part of the videos in the training data and the text description information corresponding to the videos, and the videos and thecorresponding text description information are jointly input into an action recognition model for training; the remaining training data is input into the action recognition model for training after the corresponding text description information is removed, so that effective word vectors of the text description information in the training data are trained; and the word vectors and the videos are input into a provided generative adversarial network model, and the word vectors are used as limitation conditions so that a generator in the model can generate videos.

The invention discloses a label-free cell detection device and method based on light sheet illumination. The method includes the steps that a light sheet generation unit is used for shaping a laser beam into a uniform illumination light sheet with the micron-order thickness through a cylinder lens; the light sheet excites particles or a cell suspension placed in a sample micro-cavity chamber controlled by a precise displacement table to move; a form microscopic image and a two-dimensional light scattering pattern of a single particle or cell are recorded by a detector under the focusing mode and the defocusing mode through an objective lens; and a result is input into an imaging analysis system for image processing and recognizing and classifying. By means of the light sheet illumination method, a stimulation area can be effectively limited, background interference in light scattering imaging is inhibited, and effective stimulation of the single particle or cell and particle size discrimination of the sub-micron resolution ratio level are realized; and a light sheet illumination stimulation two-dimensional light scattering technology can avoid complex dyeing operation and a fluorescence signal detection process, and label-free detection and classification are carried out on aging cells. The label-free cell detection device and method based on light sheet illumination are high in applicability and capable of being popularized.

The invention discloses a parameter identification system and method of an induction motor. The invention comprises a test system for on-line identification on a parameter of the induction motor, a reinforcement learning framework suitably used for identifying the motor parameter and a q-learning-based parameter identification method of the induction motor, wherein the test system can be used for acquiring data such as a real-time voltage, a real-time current, a real-time temperature and a real-time torque, which are needed by an on-line identification algorithm, of the motor, the data is used for identifying the parameter, the reinforcement learning framework comprises a state variable in a reinforcement learning environment, an award value and selection of an action mode, and by the q-learning-based parameter identification method, a data set is generated in real time and the parameter is identified during the running process of the test system. By the parameter identification system and method, the conflict problem of difficulty in acquiring identification accuracy and a training set is solved, a special motor mathematical model is not depended, high universality is achieved, and the data set can be generated in real time and is not needed to be prepared in advance; the identified parameter takes optimal output performance as priority and is free from influence of change of an actual physical parameter, and the identification accuracy is high; and the rotor resistance of the induction motor can be identified, and the excitation inductance of the induction motor can also be identified.

The invention relates to a scene identification method based on deconvolution deep network learning with a weight. The method comprises: (1), a weight-included deconvolution deep network model is constructed and learning is carried out on original input images by using the weight-included deconvolution deep network model, thereby obtaining feature graphs of the images on different scales; (2), feature sampling is carried out on the feature graphs obtained by learning at the step (1) by using a spatial pyramid model, thereby forming a feature vector expression of each image; and (3), the original input images are divided into a training set and a testing set and the feature vector expressions of the images are inputted into a classifier of a support vector machine respectively to complete classification training and testing, so that identification results of different scenes are obtained. According to the method provided by the invention, scene expressions with different scenes are established and thus the scene expressions become accurate and adequate, so that the scene classification precision is improved.

The invention discloses an optical remote sensing image scene classification method based on a deep twin capsule network. The method comprises the following steps of 1, deleting an average pooling layer of a deep residual network deep layer and a layer behind the average pooling layer; 2, taking the fine-tuned deep residual network as a feature extractor; 3, features of the input images are extracted respectively, and the obtained features are converted into capsule features; 4, introducing the thought of a twin network, and copying a single deep capsule network into a double deep capsule network to form two identical networks with shared feature extractor parameters; 5, calculating the distance between the two features to represent the similarity degree of the image pairs; and 6, carryingout capsule propagation by using a dynamic routing algorithm to finish image classification. Characteristic space information is stored by utilizing capsules, and a twinning network structure is combined, so that the characteristics have higher identifiability on a remote sensing data set. In addition, a regularization term is added to enhance the robustness of the model.

The invention relates to an adaptive remote sensing scene classification method based on fusion of meteorological environment parameters and image information, being able to be applied to the aspectsof geographical national condition prospecting and environment monitoring, for solving the problem that an existing method based on an image brightness value is easy to be interfered by the environment, so as to cause large feature difference based on visible light, and being not able to effectively recognize and understand the remote sensing scene. The adaptive remote sensing scene classificationmethod based on fusion of meteorological environment parameters and image information includes the steps: 1) after standardizing weather data, obtaining weather data feature Fwea by using a fully connected network; 2) using the weather data feature Fwea obtained in the step 1) to construct an adaptive convolutional neural network; 3) extracting the remote sensing image feature Frgb by using the adaptive convolutional neural network constructed in the step 2) and classifying the remote sensing image feature Frgb by means of a SoftMax classifier; and 4) training and testing the adaptive convolutional neural network, and using the trained adaptive convolutional neural network to classify the remote sensing image.

The invention discloses a method for recognizing pesticide residues on skin of cucumbers based on near infrared spectroscopic analysis and belongs to the field of pesticide residue detection. The method is characterized by comprising the following steps: a, sample preparation; b, spectral data acquisition; c, dimensionality reduction of principal component analysis spectral data; d, spectral dataclassified recognition of a supporting vector machine: S1, generating 150 sample data of a training set and 90 sample data of a testing set; S2, carrying out normalization processing; S3, selecting aGaussian radial base core function to establish a supporting vector machine model; S4, acquiring the supporting vector optimum model of the optimum parameter C and the optimum parameter g by means ofa grid search and cross verification method; and S5, recognizing the normalized training set data by applying the supporting vector machine optimum model, separately, and predicting the testing set data which are normalized. The method can be used for identifying the pesticide residues of different concentrations on skin of cucumbers in a classified manner quantitively, and is high in detection precision.

The invention relates to a method and a device for identifying a web crawler, a storage medium and electronic equipment. The method comprises the following steps: acquiring access data; determining characteristic data of the access data, the characteristic data comprising data used for representing distribution characteristics of an access interface and / or data used for representing distribution characteristics of access time; and determining that the access data is user data or crawler data according to the feature data. The method and the device are used for solving the technical problems that an anti-crawling effect based on an IP access frequency is poor and crawler recognition based on user behaviors of external interaction equipment is difficult to apply to an APP of a mobile terminal in the prior art.

The invention discloses a label-free microfluidic cell instrument and method based on a light sheet illumination and sheath flow technology, and the method is as follows: a laser beam shaping module generates a laser light sheet to be incident into a sheath flow generation control module, the sheath flow generation control module forms a sheath flow effect by a water dynamic focusing effect, the incident laser light sheet from the laser beam shaping module is coupled with a focused sample stream under the sheath flow effect to excite to-be-tested particles or cells fast flowing in the sample stream to undergo light scattering, and a two-dimensional light scattering image of the to-be-tested particles or cells is captured by an imaging acquisition module and sent to an image processing system for data analysis and result output. A sheath flow machine can be fasted formed by a novel sheath flow technology without complicated micromachining operation; by the coupling of the light sheet illumination technology and the sheath flow technology, the light scattering detection signal to noise ratio can be improved, and label-free fast imaging of the flowing particles or cells can be realized. High-precision particle size recognition in a flow state can be realized by a particle size identification method based on Euclidean distance for matching degree query.

The embodiment of the invention discloses a network content recall method and device, and the method comprises the steps of determining a long-term feature group and a recent feature group from the historical browsing data according to the generation time of the elements in the historical browsing data of a user; and determining a long-term coding vector corresponding to the long-term feature group according to the similarity between elements in the long-term feature group; according to the similarity degree between the long-term coding vector and the elements in the recent feature group, obtaining the long-term coding vector, determining a user feature vector corresponding to the recent feature group, therefore, so that the information in accordance with the interest preference browsed by the user in the user feature vector can be highlighted, and the other information is retained to a certain extent. The recommendation candidate item determined by the user feature vector better conforms to the actual browsing interest of the user, and has certain generalization, so that the quality of network content recommendation for the user is improved, and the browsing experience of the user is improved.

The embodiment of the invention provides a training method, a detection method and a system for a malicious webpage detection model. The training method comprises the following steps of: obtaining a data set of a webpage and judgment result data of whether the webpage is a malicious webpage or not; Processing the data set; Establishing a training model architecture; And taking the processing result of the webpage data set as input data, and taking the judgment result as output data to train the training model architecture, so as to form a detection model capable of predicting whether the webpage is a malicious webpage based on the input webpage data processing result. According to the training method in the embodiment of the invention, the detection model for simply and efficiently predicting whether the corresponding webpage is the malicious webpage based on the input network data can be trained, so that the precision is higher when the webpage is predicted through the detection model, and meanwhile, the prediction operation of a user on the webpage is greatly simplified.

The invention discloses a webpage text classification method based on an enhanced capsule network and a storage medium thereof, and the method comprises the steps: crawling webpage text data in a specific field, carrying out the cleaning and data structuralization of the obtained text data, and finally obtaining an experiment corpus; setting an architecture of an enhanced capsule network, whereinthe architecture sequentially comprises a dense convolutional network, a main capsule layer and a digital capsule layer; and training the enhanced capsule network by taking the training data in the training set as the input of the enhanced capsule network to obtain a classifier, and verifying the accuracy of the classifier by using the test data of the test set. According to the method, the denseconvolutional network is introduced to extract the feature information, so that the features are more discriminative, and the learning ability of the model on a data set is improved. And the main capsule layer is further encoded by adopting a dynamic routing mechanism, so that the obtained features are more directional, and the capsule network is more robust.

The invention discloses a crowd fleeing event detection method and system based on a direction-clustering model. In order to solve the problem that the crowd fleeing event detection method is low in detection accuracy, the technical scheme comprises the steps of extracting moving tracks from crowd monitoring videos, thereby obtaining a track coordinate matrix, and calculating speeds and directions of moving particles on the moving tracks through the track coordinate matrix; on the basis of the randomness of the speeds and directions of the moving particles, establishing the direction-clustering model which represents a moving state of a crowd; calculating a direction-clustering index of each frame of crowd monitoring video according to the direction-clustering model; extracting an event detection threshold self-adaptively from all obtained direction-cluster indexes, comparing the event detection threshold with the direction-clustering index of each frame of crowd monitoring video, and finally detecting a crowd fleeing event of the crowd monitoring video. The method is widely applied to a middle and high density crowd and the detection accuracy is high.

The invention discloses a social media named entity identification method based on affix perception. The method comprises the following steps: collecting a social media data set marked with named entities; capturing embedded representation, character level representation and affix feature representation of the words, and fusing the embedded representation, the character level representation and the affix feature representation of the words to serve as final representation of the words; inputting the obtained final representation of the word into a bidirectional convolutional neural network anda conditional random field, predicting a label sequence and calculating a loss value; training the model by adopting a random gradient descent algorithm according to the obtained loss value; and inputting the text into the trained model, and identifying the named entity in the text. According to the method, semantic representation of the words is enriched, the problem of unregistered words in thesocial media data is relieved, and the named entity recognition effect is improved.

The invention belongs to the technical field of machine learning and intelligent human-computer interaction, and discloses an emotion-recognition-oriented electroencephalogram signal channel selectionmethod, system and application. The method comprises the following steps: performing debasing preprocessing on EEG data; calculating the power spectrum intensity of a frequency domain signal by combining a sliding window and fast Fourier transform, and taking the power spectrum intensity as electroencephalogram characteristics; respectively solving the weight of each feature by adopting ReliefF and MIC algorithms, performing integrating by utilizing a wave arrival counting method to obtain an integral sum of each channel, sequentially adding feature data of a channel with a relatively large integral sum value, performing classifying by adopting a random forest, and finding out an optimal channel subset; and performing classification evaluation. According to the method, a feature selectionmethod combining ReliefF and MIC algorithms is adopted, each channel serves as a whole, the purpose of greatly reducing the number of the channels is achieved, the efficiency of the system can be improved, the real-time performance of the system can be improved, and the method has important significance in the fields of electroencephalogram emotion recognition and intelligent man-machine interaction.

The invention provides a hydrogen-containing energy storage microgrid optimization operation method based on deep reinforcement learning, and the method comprises the steps: building an electrolytic cell efficiency characteristic model through a linear interpolation method, combining the electrolytic cell efficiency characteristic model with a gas turbine model, an electrochemical energy storage model, a hydrogen storage tank model, a fuel cell model and the like, and taking the microgrid operation cost as a target, constructing an optimized operation model of the hydrogen-containing energy storage micro-grid. And finally, solving a sequence decision problem of micro-grid optimization operation by adopting a depth deterministic strategy gradient algorithm. According to the method, the efficiency characteristic of the electrolytic cell is considered, the hydrogen energy storage capacity can be fully utilized, the optimization problem is solved according to the deep reinforcement learning principle, the operation cost of the hydrogen-containing energy storage microgrid is reduced, and good generalization is achieved.

The invention discloses a method and a device for inverting a visible image of a remote sensing image in a bare place. The method comprises the following steps: preprocessing original remote sensing data to obtain a test image of a target geographic area and an RGB image of the target geographic area; performing pixel-level semantic segmentation on the test image of the target geographic area by using the first classification network model; performing region contour selection on the test image after semantic segmentation by using a second classification network model; and aligning and superposing the selected region contour to a corresponding position on the RGB image of the target geographic region. The device comprises a preprocessing unit, a semantic segmentation unit, a contour selection unit and a contour output unit. The method and device fully automatically draw the bare land inversion graph, save a large amount of manpower, have generalization, and can be made into an interface to be popularized to various cities. Full-vector diagram drawing is carried out, the definition of contours and numbers cannot be influenced by zooming in and zooming out of the drawn diagram, and the diagram is clearer than an Arcgis manually drawn diagram.

The invention discloses a smart street lamp network system based on the blockchain technology. The smart street lamp network system comprises a smart street lamp, and a single lamp controller, a loopcontroller, a private chain background server and an Ethereum bottom layer system, which are arranged in the smart street lamp; the single lamp controller and the loop controller are arranged in the smart street lamp and perform data exchange through the private chain background server; and the Ethereum bottom layer system comprises network data layer, a contract consensus layer and an applicationlayer, the single lamp controller is located on the application layer, the loop controller is located on the contract consensus layer, and the private chain background server is located on the network data layer. By means of this manner, the embedded operation mode of the existing street lamp box can be converted into a mode of transferring the data calculation pressure and the storage pressure to a cloud platform through a reliable data transmission mode; the characteristics of high data security and accuracy and high traceability of the Ethereum bottom layer system can be effectively combined; and the smart street lamp network system has the characteristics of high scalability, generalization and high practicability.

The invention discloses a photovoltaic panel dust detection system and method based on an unmanned aerial vehicle. The system comprises a carrying platform which is provided with an image collection device and a background server. The image collection device is in signal connection with the background server through a communication network. An operation control program is set in the carrying platform, and the image collection device is a pan-tilt camera; and an image processing model is arranged in the background server. The carrying platform is a multi-rotor unmanned aerial vehicle; the system comprises a tablet device, and a route planning APP is arranged in the tablet device. An RTK module is arranged on the multi-rotor unmanned aerial vehicle, the multi-rotor unmanned aerial vehicle isconnected to a thought network through the RTK module, and the positioning precision of the equipment is improved. By the adoption of the technical scheme, dust on the solar cell panel can be automatically detected, and whether cleaning is needed or not can be judged.

The invention discloses a design method for deploying and an optimizing operator library on an FPGA and a DSP. The method comprises the following steps: designing an underlying hardware operator library corresponding to an operator library of a high-level deep learning framework, and abstractly packaging operators in a lightweight network to form a fusion operator library; packaging the fusion operator library into a parallel operator library with hardware characteristics by adopting a preset segmentation strategy according to the fusion operator library in combination with computing resources of hardware; and combining the parallel operator library with a rearrangement strategy. According to the method, technical support is provided for quickly completing deployment and optimization of the deep learning network on resource-limited side-end equipment such as the DSP and the FPGA. The core of the method is to construct an underlying deep learning operator library with high practicability and high mobility. The operator library is combined with hardware characteristics and fused with efficient strategies such as heuristic segmentation and data flow rearrangement. And the basic requirements of the neural network model on the deployment of the FPGA and the multi-core DSP can be met.

The invention discloses a training task generation method of an upper limb rehabilitation robot based on impedance variable demonstration learning. Diversified training tasks are generated through thesteps of man-machine interaction model building, TP-GMM learning, local rigidity estimation, trajectory regression, task generalization and the like in sequence. According to the method, the trajectory, interaction force and impedance information of the task are comprehensively considered, so that the training task generated by the method not only simulates the movement behavior of a therapist / technician, but also simulates the impedance variable strategy in the interaction between the therapist / technician and the upper limb rehabilitation robot. According to the invention, the TP-GMM modeling method is applied to coding of the training task of the upper limb rehabilitation robot for the first time, and the generalization function of the variable starting point / target point of the task can be realized. According to the invention, the problem of complex programming required by diversification of training tasks is solved, boring and tedious workload of adjusting task parameters for a patient in real time by the therapist / technician can be reduced, and rehabilitation training efficiency is improved.

The embodiment of the invention provides a training method of a user intention recognition model, a server and a display device, the training method comprises the steps that samples in a training sample set are divided into a plurality of training groups, and each training group comprises a support set and a query set; inputting the training group into a user intention recognition model, and respectively extracting sample features of a support set and sample features of a query set; calculating the similarity between the sample features of the support set and the sample features of the query set through a cosine function; calculating a loss function of the user intention recognition model according to the similarity; and optimizing the loss function through the plurality of training groups until the user intention recognition model converges. The user intention recognition model is trained based on a small sample classification mode, sample data can be fully utilized, and generalization and recognition accuracy of the user intention recognition model are improved.

The embodiment of the invention discloses a signal detection method, a signal detection device, computer equipment and a storage medium. The method comprises the steps: receiving an original signal which comprises at least two sub-signals; determining the attention between the sub-signals; determining signal characteristics representing periodicity of the sub-signals according to the attention; and determining sub-signals in an abnormal state according to the signal characteristics. The attention among the sub-signals has generalization; according to the invention. The attention is not required to be paid to the signal type of the original signal and attention is only required to be paid to during feature extraction, so that anomaly detection can be performed on different types of originalsignals through attention. The generalization performance and adaptive capability of anomaly detection are improved, re-signal feature and data distribution establishment for different types of signals are not required. The workload is reduced.

The invention provides a multi-speaker speech synthesis method based on probability generation and a non-autoregression model. The method comprises the steps that S1, a speaker personalized encoder and a probability generation encoder receive a target Mel spectrum and extract a speaker personalized vector and a probability generation vector respectively; s2, encoding the spliced and fused input vectors based on an encoder of a deep network to obtain phoneme-level deep features; s3, a phoneme duration predictor receives the spliced and fused fusion features and predicts a phoneme duration sequence; s4, receiving the phoneme time length sequence and expanding the fusion feature by the length regulation network to obtain a frame-level feature; s5, a decoder based on the deep network receives the frame-level features and maps the frame-level features into a predicted Mel spectrum, and a post-processing network supplements residual information of the predicted Mel spectrum; and S6, the vocoder maps the predicted Mel spectrum supplemented with the residual information into a sound waveform to obtain synthetic speech. The generalization of a multi-speaker speech synthesis system and the similarity of synthesized speech can be improved.

The invention provides a drug-target protein affinity prediction method and system, and belongs to the technical field of biological information processing based on artificial intelligence. The method comprises the following steps: respectively coding a drug chemical molecule and a target protein molecule in a drug-target protein pair to be detected to obtain a drug input representation and a target protein input representation ; and processing the drug input representation and the target protein input representation by using a drug-target protein affinity prediction model to obtain a drug-target protein affinity prediction value. A meta-learning algorithm is utilized to effectively mine the correlation between different subtasks; and the task unevenness between different subtasks is relieved by using the regularization item, the task adaptability of the drug-target protein affinity prediction model is improved, the drug-target protein affinity prediction model with generalization performance is obtained, the accuracy of unknown drug-target protein pair affinity prediction is improved, and the cold start problem in the aspect of drug-target protein affinity prediction is effectively solved.

The invention discloses a multi-feature fusion fatigue detection method based on deep learning and machine learning. The method comprises the following steps: S1, data acquisition: collecting a fatigue face image; S2, constructing an expression recognition data set; S3, extracting an attention feature map: inputting the expression recognition data set into a deep residual network to obtain the attention feature map, then adding the attention feature map into a newly constructed new data set containing fatigue facial expressions, and constructing a data set of the attention feature map; S4, inputting the expression recognition data set containing the attention feature map into a 19-layer convolutional neural network VGG19 for training; S5, extracting traditional fatigue features; S6, extracting a deep learning confidence coefficient; and S7, fusing multiple features to train a machine learning classifier. According to the method, based on the expression recognition model guided by the attention feature map, the deep learning network is used, the attention is focused on the eye and mouth areas with the most abundant features on the face, and the recognition precision of expression recognition can be improved.

The invention relates to a food image segmentation method and system based on dynamic transformer, and the method comprises the steps: S1, dividing an input food image into a series of image blocks with different sizes, and inputting the image blocks into a plurality of dynamic visual transformer encoder networks with different sizes; outputting multi-layer image feature vectors of different scales; s2, extracting the image feature vectors of the preset layer for fusion to obtain fused image feature vectors; s3, constructing a multi-stage feature aggregation network, performing top-down feature fusion on the fused image feature vectors, constructing a feature pyramid, and obtaining multi-scale feature fusion vectors; and S4, constructing a segmentation decoder, performing convolution and up-sampling operation on the multi-scale features fused by the feature pyramid, and finally generating a segmentation result with accurate food category boundary segmentation. The method provided by the invention can be adaptive to different picture scales, and richness and integrity of picture semantic information extraction are improved, so that a food segmentation model has generalization and robustness.

The invention provides a commodity information recommendation method and device and a storage medium, and the method comprises the steps: importing a commodity information data set, and dividing the data set of the commodity information data set to obtain a commodity information training set and a commodity information test set; performing vectorization analysis on the commodity information training set to obtain a knowledge graph matrix and a training set matrix. According to the method, implicit characteristics among data can be effectively mined, accurate recommendation is realized, better differentiation can be realized in mathematics, more effective information in a graph can be mined, rapid recommendation can be realized, accurate recommendation can be realized in a specific scene, certain generalization is realized, the method has a certain effect on different types of data, can effectively improve the recommendation accuracy, has good robustness, can reasonably and accurately recommend commodities to the user, realizes cold start commodity recommendation to the user, and improves the recommendation reliability and accuracy.

The invention discloses a robot autonomous anomaly repair skill learning method and system. The method comprises the following steps: pre-defining a human demonstration track when a robot executes a complex task; acquiring multi-modal sensing information of the robot when executing the complex task, and monitoring by utilizing the multi-modal sensing information to acquire an abnormal motion skill; and based on the abnormal state type of the abnormal motion skill, starting an adaptive motion repair strategy to autonomously repair the abnormal motion skill. In the embodiment of the invention, the abnormal types encountered when the robot executes the complex task can be distinguished, and the corresponding repair strategy can be formulated, so that the robot can be promoted to realize longer-term autonomous operation.