244 results about "Web learning" patented technology

Generalized learning rules may be implemented. A framework may be used to enable adaptive spiking neuron signal processing system to flexibly combine different learning rules (supervised, unsupervised, reinforcement learning) with different methods (online or batch learning). The generalized learning framework may employ time-averaged performance function as the learning measure thereby enabling modular architecture where learning tasks are separated from control tasks, so that changes in one of the modules do not necessitate changes within the other. Separation of learning tasks from the control tasks implementations may allow dynamic reconfiguration of the learning block in response to a task change or learning method change in real time. The generalized spiking neuron learning apparatus may be capable of implementing several learning rules concurrently based on the desired control application and without requiring users to explicitly identify the required learning rule composition for that task.

A 1D-CNN-based ((one-dimensional convolutional neural network)-based) distributed optical fiber sensing signal feature learning and classification method is provided, which solves a problem that an existing distributed optical fiber sensing system has poor adaptive ability to a complex and changing environment and consumes time and effort due to adoption of manually extracted distinguishable event features, The method includes steps of: segmenting time sequences of distributed optical fiber sensing acoustic and vibration signals acquired at all spatial points, and building a typical event signal dataset; constructing a 1D-CNN model, conducting iterative update training of the network through typical event signals in a training dataset to obtain optimal network parameters, and learning and extracting 1D-CNN distinguishable features of different types of events through an optimal network to obtain typical event signal feature sets; and after training different types of classifiers through the typical event signal feature sets, screening out an optimal classifier.

The invention relates to an image segmentation method based on a multi-supervision full-convolution neural network. According to the invention, further improvement is carried out based on the full convolution neural network, so a novel network structure is provided. The network structure has three edge output layers with supervision which are capable of guiding a network to learn multi-scale features and allowing the network to acquire local features and global features of images. Meanwhile, in order to keep context information in the images, in the upper sampling parts of the network, upper sampling is performed on output feature images by sue of multiple feature channels. Finally, a fusion layer with the weight is used for fusing the classification results of the multiple edge output layers, so the final image segmentation result is obtained. The method is characterized by high segmentation precision and quick segmentation speed. In osteosarcoma CT data segmentation, the DSC coefficient of the acquired segmentation result by use of the provided algorithm reaches 86.88%, which is higher than that of a traditional FCN algorithm.

A system controlling online access to a study course verifies the identity of an individual taking a study course over a global computer network from a first computer at a node of the network. The first computer has a biometric identification program and communicates over the network with a second computer that is at a network node other than a node of the first computer. The second computer includes study course material. The first computer operates a biometric reader, which obtains a first set of biometric data from the individual and a second set of biometric data from the individual while access is granted to course material. The biometric identification program compares the first set of data with the second set of data to make a verification of the identity of the individual and communicates the verification to the second computer.

A method and apparatus for automatically performing medical image analysis tasks using deep image-to-image network (DI2IN) learning. An input medical image of a patient is received. An output image that provides a result of a target medical image analysis task on the input medical image is automatically generated using a trained deep image-to-image network (DI2IN). The trained DI2IN uses a conditional random field (CRF) energy function to estimate the output image based on the input medical image and uses a trained deep learning network to model unary and pairwise terms of the CRF energy function. The DI2IN may be trained on an image with multiple resolutions. The input image may be split into multiple parts and a separate DI2IN may be trained for each part. Furthermore, the multi-scale and multi-part schemes can be combined to train a multi-scale multi-part DI2IN.

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 a system and a method for recognizing a remote sensing image target based on migration network learning, mainly solving the problems that the correct recognition rate for a remote sensing image with a label is relatively low when the number of data is less and the obtaining of the image label is difficult and needs high cost in the conventional methods. The whole system comprises an image characteristic extracting module, a migration network classifier learning system generating module and a migration network classifier learning system learning module, wherein the image characteristic extracting module is used for completing the characteristic extraction of the image; the migration network classifier learning system generating module is used for training input sample data by a network integrated learning algorithm introduced into migration learning to obtain a migration network classifier learning system; and the migration network classifier learning system learning module is used for completing the classification and the recognition of the characteristics of a new sample image. The invention has the advantage of the capability of utilizing other existing resources to improve the correct recognition rate of the remote sensing image target without collecting data again and can be used for the target recognition of the remote sensing image.

The invention discloses an image super-resolution enhancement method based on a bidirectional recursion convolution nerve network. The method comprises two parts, i.e., model training and resolution enhancement. The model training comprises: obtaining multiple groups of high and low-resolution images comprising the completely same content, extracting high and low-resolution image block sequences through a window sliding method, and accordingly, training a bidirectional recursion convolution nerve network model. The resolution enhancement comprises: segmenting low-resolution images to be processed into a group of low-resolution image blocks to be processed through the window sliding method, generating corresponding high-resolution image blocks through the well trained bidirectional recursion convolution nerve network model, and generating corresponding high-resolution images through fusion. According to the method, space position relations of adjacent image blocks are learnt by full use of the bidirectional recursion convolution nerve network, and accordingly, the high-resolution images comprising detail information such as more edge textures and the like are obtained.

The invention belongs to the technical field of computer vision, discloses an image semantic segmentation method based on confrontation training and is used for solving a problem that the existing semantic segmentation method cannot achieve meaningful fusion of global information and local information of images and cannot perform learning on high-order potential energy in the images. According tothe invention, a loss function of the whole network is defined based on a confrontation training network to serve as a confrontation network of a general function approximator, thereby not only beingcapable of learning how to combine different levels of information, but also being capable of forcing the generated network to learn information such as single points, paring and high-order potentialenergy in a segmented picture, achieving organic integration of local features and global features of the image and acquiring a segmentation image with the effect being more authentic; and meanwhile,the layer-by-layer training method avoids the complex network parameter initialization process, and enabling the whole network to use a method of random initialization.

The invention discloses a rotating machinery fault diagnosis method based on a one-dimensional depth residual convolution neural network. Firstly, the network learns more deep and abstract fault characteristics of training samples through a stacked one-dimensional residual module. Then, the Adam optimization algorithm is used to optimize all the super-parameters to complete the deep-level featureextraction and fault classification, and the rotating machinery fault diagnosis system model based on one-dimensional depth residual convolution neural network is obtained. Finally, the test samples are inputted into the trained fault diagnosis model, and the deep-seated features are extracted automatically to diagnose the health status of the rotating machinery.

The invention discloses an image object counting method based on a convolutional neural network. The image object counting method makes the characteristics learned by the network robust to the targetdeformation through a robust enhancement layer, and lowers the computational complexity of the model; a pyramid layered counting module is used to perform density estimation, and the multi-scale information contained in hierarchical features of the convolutional neural network are fully used, which significantly improves the computational efficiency while accurate counting is achieved. In summary,the image object counting method is based on the convolutional neural network, achieves accurate counting of the target in an image, and can be applied to a target counting task in a complex scene, and is low in computational complexity and high in practical application value.

The invention discloses a feedforward neural network hardware realization method based on multicore technology, relating to a computer system based on a specific calculation model. The invention adopts an NIOS II soft core processor and utilizes Quartus II software in an FPGA development system thereof to build a multicore processor system, the number of processor cores is only limited by chip scale, each processor core can be independently programmed, and multiple processor cores can simultaneously run respective program and simulate the characteristics of nerve cells of any kind by performing programming of one single core; in the network studying process, the single cores calculate output amount in a parallel mode and finish the adjustment of corresponding weights and thresholds; the input and output of each single core are performed with data exchange via a multiport memory to stimulate feedforward neural network with any structure, which realizes the purposes of high speed, distribution and parallel calculation of the neural network.

The invention discloses a binocular stereo matching method and a binocular stereo matching system based on dense network depth learning. Firstly, stereoscopic image pairs with real values are preprocessed to construct training samples for network learning. Then a dense convolution neural network model is trained to calculate the matching cost. Finally, cross-aggregation algorithm is used to optimize the matching cost, and the optimal disparity is calculated by WTA strategy, and the corresponding disparity map is obtained. A method for extracting features of convolution layer features that a dense convolution neural network model is used to extract features of convolution lay, Attempt to find a better matching cost, effectively solve the problem that pixel matching point can not be found accurately in weak texture region and the performance of detail features is poor, with higher computational accuracy and better robustness.

The invention provides an image pixel classifying method based on the convolutional neural network. The method mainly includes the steps of image inputting, self-help network processing, iterative processing, image thinning and estimation result obtaining. The process includes estimating depth and camera motion by adopting the convolutional network and includes three stages that image pairs are sampled from a scene data set, and the image pairs with highlight conformity errors are discarded; the image pairs obtained after preprocessing are input into a self-help network to calculate optical flow, depth and camera motion; an iterative network is used for conducting iterative repetition multiple times to improve an existing estimation result; a high-resolution depth image and motion estimation are obtained after refining of a refining network. The network is obviously superior to a traditional motion structure, and the result is more accurate and robuster; different from the network of estimating depth through the single image, the network learns the concept of learning, is capable of using motion parallax for processing a new type of scene and allows motion estimation.

The invention relates to a method for realizing super resolution for an image and belongs to the computer vision field. The method includes the following steps of: A1, data preprocessing: a certain number of high-resolution natural images are adopted to form a data set, a certain number of image blocks are extracted from the data set, Bicubic interpolation down-sampling and up-sampling in three times are carried out on the image blocks, and low-resolution images can be obtained; A2, network structure design: a designed convolutional neural network has 4 layers altogether; A3, hyper parameter selection: parameters such as network learning rate, learning momentum and batch_size are determined; and A4, network training and super parameter optimization: the convolutional neural network of all images in the training set from low-resolution images to corresponding high-resolution images is trained, and after any one image is inputted into the trained network, a high-resolution image can be obtained, so that the super resolution of the image can be realized.

The invention relates to a network curriculum learning platform and a communication method. The network curriculum learning platform comprises a WEB server, a logic server, a database server, a content integration server, a content publishing server, a material server, a streaming media server and a network curriculum packet server, wherein the network curriculum packet server stores a network learning content packet. The network curriculum learning platform is characterized in that: the network learning content packet and the WEB server perform interactive communication through an application program interface; and the network curriculum learning platform reads basic information of a user to the network learning content packet through the application program interface on the WEB server and receives dynamic learning information of the user transmitted by the network learning content packet. By the networking learning platform and the communication method, the modularization degree is high; the network curriculum is easy and convenient to transplant; and learning resources can be shared highly.

The invention discloses an increment neural network and sub-graph code based image classification method. The increment neural network and sub-graph code based image classification method comprises the following steps of: extracting local characteristics; studying increment code books in a network; performing characteristic coding based on a sub-graph; getting image spaces together; and studying a classifier and predicting a model. The increment neural network and sub-graph code based image classification method greatly reduces time complexity of a traditional algorithm to a great extent as the code book can be studied efficiently, and the space relationship between visual words is also kept; in addition, the characteristic coding based on the sub-graph is specifically for performing characteristic coding by fully utilizing the space relationship between the visual words, so that more abundant semantic information can be extracted, and excellent classification performances are obtained finally while the computational efficiency of a classification system is increased. Therefore, the increment neural network and sub-graph code based image classification method has a relatively high use value.

The invention provides a personalized recommendation system and a personalized recommendation method based on a deep neural network. The method includes: by fusing the candidate set generation moduleof the depth neural network, combining multi-user and multi-item features to conduct deep neural network learning to generate candidate sets, and based on a ranking set generation module fusing the deep neural network, combining the multi-user and multi-item features to conduct the depth neural network learning and scoring on the candidate set, to generate a better ranking set containing the personalized recommendation content, and finally, based on the collaborative filtering algorithm and the ranking set, carrying out further personalized recommendation, and obtaining the final recommendation list. Through the above way, the multi-user and multi-item are combined to improve the accuracy of a recommendation process, and accurate personalized recommendation is achieved in combination witha collaborative filtering algorithm. As a result, the personalized recommendation efficiency and user experience are improved.

The invention discloses an online learning monitoring feedback system, which comprises a database server, an application server, a computer terminal and a control console. The data server, the application server and the computer terminal are respectively connected with the control console. The data server is interconnected with the application server. The database server is provided with a data management module for managing a learner online learning behavior information base, a learner information base and a course knowledge point structure information base. The application server is provided with a learning behavior record module, an information statistical analysis module, and a learning decision support module. The computer terminal is provided with a subsidiary effective learning feedback information module, a course design feedback information module and a learning content design feedback information module. By adopting the system, not only the time utilization ratio of learners and the learning efficiency are improved, but also scientific proposals and scientific testing methods are provided for designing and improving online courses for online course designers.

The invention discloses a session social recommendation method based on context neighborhood modeling. According to the method, based on historical interaction data and social network information of agiven target user, the next item most likely to be interacted by the target user is found out. According to the method, firstly, user interests are modeled, and then the representation of each friendin a social network corresponding to a target user is obtained by adopting a session-level attention mechanism. And the social influence of friends on the user is learned by using a social network. And finally, article recommendation is performed in combination with the social influence of friends on the target user and the interest of the target user. The method overcomes the defects in the existing method that the interests between the target user and friends are ignored and only partially coincide. Therefore, compared with an existing method, the recommendation effect implemented by the method is remarkably improved.

The invention relates to the field of human body behavior recognition, in particular to an angle independence-based skeleton behavior recognition method, system and device, and aims to improve the accuracy of angle-independent skeleton behavior recognition. The angle independence-based skeleton behavior recognition method comprises the steps of designing a specific visual angle sub-network on thebasis of a skeleton sequence of each visual angle, focusing on key joint points and key frames through space domain attention and time domain attention modules respectively, and learning a discrimination characteristic of each visual angle sequence through a multi-layer long-short-term memory network; serially connecting output characteristics of all the specific visual angle sub-networks to serveas an input of a public sub-network, further learning angle-independent characteristics through a bidirectional long-short-term memory network, and focusing on a key visual angle through a visual angle attention module; and proposing a regularization cross entropy loss function for promoting the modules of the network to jointly learn. According to the skeleton behavior recognition method, systemand device, the recognition accuracy is effectively improved, and visual angle characteristics with relatively numerous learning information can be automatically focused.

The invention discloses a method for obtaining user specific metaknowledge interest. The method is specific to a learner in a network learning system, aims at exactly identifying and obtaining learning interest of the network learner, predicts the future interest change of the learner, and improves learning efficiency and performance of the network learner. In the aspect of user model, a user interest model facing the network learner group is adopted. In the interest omission algorithm of dynamic interest perception, an interest omission algorithm featuring automatically discovering and omitting the threshold is adopted. In the interest prediction algorithm, a prediction method featuring user collaborative prediction is adopted. The invention analyzes blogs of the learner to exactly identify the interest change of the user towards metaknowledge, conception and learning resource, reflects interest degree in a quantitative way, predicts the interest change of the learner, improves defects of the existing technologies, and has good application prospect in the network learning system and network individual service.

The invention discloses a word embedding-based deep zero sample classification method. Types not appearing in a learning stage still can be identified in a prediction stage. A word embedding and depthfeature extraction-based zero sample image learning model is proposed; according to the method, a subspace of image feature and semantic feature joint embedding is learnt through a strong deep neuralnetwork; and by utilizing a semantic capability of word vectors, the purpose of zero sample learning of an image is achieved. A deep learning method and zero sample learning combination-based model is proposed; a large amount of improvements are made for a combination part embedding space; reference and guidance are provided for learning of image zero samples or zero samples in other modes; and the shortcomings of indefinite sample attribute definition, manual feature extraction and the like in conventional zero sample learning are overcome. The method is widely used in an end-to-end type prediction task, and is especially suitable for the type prediction task short of training samples and even short of a type of training samples.

The invention discloses a method for learning driving styles based on a self-coded regularization network. The method mainly comprises the following steps: performing GPS (Global Positioning System) data conversion, performing regularization network self coding, performing target function sum approximation, establishing a run length encoding frame, and establishing the number of drivers, namely, in a group of unknown driving, inputting GPS data of vehicles establishing a statistic characteristic matrix as network input, introducing a marker of a limited training set as a prior into an unsupervised automatic encoder, reconstructing hidden layer RNN (Recurrent Neural Network) characteristics, extracting a neck layer of a regularization self-coding structure as a final driving style characteristic representation layer, and estimating the number of drivers in the driving process. By adopting the method, the limit that the driving style of an unknown driver is hard to describe can be solved, a self-coded regularization network is designed to directly learn driving habits of the driver from the GPS data, then recognition and classification precision of different drivers can be improved, and a relatively safe and accurate method can be provided for design of assistant and automatic driving systems.

The invention discloses a method for counting persons in a video based on LSTM (Long Short-Term Memory)-weighted neural network, and the method comprises the steps: firstly estimating a perspective according to different scenarios, and then generating an adaptive density map of the crowd; carrying out the down-sampling of continuous multi-frame images, inputting the images to a neural network, andtraining the network based on the images and its corresponding real adaptive density map; estimating the density map of the input images through the trained network, and predicting the number of persons based on the density map. According to the scale difference of the crowd in the scenarios, the method measures the different scale features obtained through network learning through a scale loss function. The method employs a weighted loss function for balancing the contributions of different regions for solving a problem that the crowd is not uniform in distribution. Meanwhile, the method obtains the correlation information of the adjacent frames through LSTM, performs the postprocessing of the predicted number of persons through smooth filtering, and improves the person counting accuracy.

The invention discloses a creating and using method of a multi-user network interdynamic learning system for dissolving the problem that the student learns and acquires the enough knowledge contents under the limit time and energy, which comprises the following steps: (A) establishing a learning data base at the server and inputting the classified learning data and the corresponding learning method file; (B) acquiring the learning account number by registering at the server; (C) analyzing the data accessing and the learning effect state of the learner by the learning system when the learner learns on the internet, generating the special statistical data, saving in the statistical table of the learning state, and estimating the learning data and the learning method of the data base automatically according to the statistical data; (D) providing the statistical data and the point rating of the original learners when the new learner learns and directing the new learner to learn. The invention makes thousands of people in different sections and different estates learn on the internet, which acquires the best learning effect with the lowest cost.

The invention discloses a video pedestrian re-identification method and system based on pedestrian average status. The method includes: shooting and obtaining a video by non-overlapping cameras to form a data set containing multiple pedestrians, training a Siamese network with a Siamese loss function together, and due to that video features learned by a trained network have distinguishability, initializing average bodies of individuals through feature vectors obtained by averaging out features of all frames from the same video; defining an in-video loss function as minimizing distance betweenthe video features and the average status, combining the in-video loss function and the Siamese loss function together, and starting from beginning to train a new model; and using the trained model toupdate the pedestrian average status, and using the updated average status to retrain the network until further gain on accuracy of pedestrian re-identification is not generated on a validation set.The method of the application is superior to most of most advanced algorithms on PRID 2011, iLIDS-VID and MARS data sets, and thus effectiveness of the method provided by the application is also proven.

The invention relates to a photovoltaic power generation forecasting method based on a GRNN. According to the method, firstly factors such as solar terms, weather, and sunlight are considered to establish a photovoltaic curve pattern, then a photovoltaic power generation forecasting model based on the GRNN is provided, and a solution algorithm design is conducted. The method has the following advantages that weight does not need to be amended through error reverse calculation in the training process of the GRNN, and the transfer function can be adjusted by just changing a smoothing parameter Sigma, so that the training time is reduced, and the network learning speed is increased; a GRNN forecasting model is high in nonlinear mapping ability and good in approximation performance, and has good robustness, and thus is suitable for processing unstable data; the photovoltaic power generation forecasting technology of the GRNN obviously improves the forecasting accuracy; the forecasting result can provides decision information for grid photovoltaic scheduling, and has great significance in guaranteeing safe operation of the grid.