74 results about "Recognition memory" patented technology

The invention discloses a generation method of an image description from a structured text. The generation method comprises the steps of downloading pictures from the internet to form a picture training set; conducting morphological analysis on descriptions which correspond to the pictures in the picture training set to form the structured text; using an existing neural network model to extract convolution neural network characteristics of the pictures in the training set, and using <, picture characteristics and structured text < as inputs to form a multitasking recognition model; using the structured text extracted from the training set and a description which corresponds to the structured text as inputs of a recurrent neural network, and conducting training to obtain a parameter of a recurrent neural network model; inputting the convolution neural network characteristics of an image ready to be described, and obtaining a predicted structured text through the multitasking recognition model; inputting the predicted structured text, and obtaining the image description through the recurrent neural network model. Compared with the prior art, a better image description effect, accuracy and sentence variety can be generated through the method, and the generation method of the image description from the structured text can be effectively popularized in an application of image retrieval.

The invention provides a recognition method of named entities in microblog messages. The recognition method includes that a few named entities are specified as seeds; a certain number of microblog messages from the original microblog message set to be processed are automatically marked as a training data set; and then the training data set is utilized to train a named entity identifier and the trained named entity recognizer is utilized to recognize the named entities in the microblog messages. According to the recognition method of the named entities in the microblog messages, only a few existing seed entities need to be specified to enable a high quality training set to be automatically marked; the labor costs are significantly reduced for the microblog messages which are texts capable of being updated rapidly; and an iterative mode is utilized to generate high quality marked data step by step and each time first N newly named entities which can most reflect the appearing law of the named entities in real microblog data are selected to add into a seed bank, so that finally generated marked data can well cover the integral microblog message set.

The invention discloses a film comment-oriented naming entity recognition method, and belongs to the technical field of entity recognition. The method comprises the following steps of: defining a naming entity type of a name in a film comment; constructing a training data set TrainSet; segmenting a non-marked data set by taking characters as units so as to construct a character matrix, and constructing a mapping dictionary of characters and word vectors by utilizing a skip-gram model in word2vec; generating a naming feature dictionary corresponding to each character; combining word vectors andnaming features of the characters to generate a feature vector matrix for training a name recognition model; constructing a mixed model by utilizing Bi-LSTM and CRF so as to train the name recognition model; and carrying out name recognition on the feature vector matrix through the trained name recognition model, so as to realized film comment name recognition task-oriented name definition, carryout name labeling on character levels during name recognition, obtain name features and combine word vectors obtained via word2vec training with the naming features.

The invention provides a conditional random field-based automatic Chinese personal name recognition method. An automatic Chinese personal name recognition system is constructed by studying Chinese personal name characteristics and combining a statistical probability model; text messages are subjected to segmentation, the conditional random field is based, a context rule and credibility method is combined, and candidate personal names are obtained; personal names with boundary recognition errors are corrected by means of a local statistical algorithm, and recognition results of a system are finally obtained. The recognition errors generated through segmentation are greatly reduced, the problem that other named entities are recognized as Chinese personal names is better solved, and the recognition effect is improved.

The invention puts forward a recognition method of key nodes of a social network in a topic field with the aim of solving the problem in the prior art that influence factors of user's nodes on information dissemination in a real social network are not taken into consideration and computing complexity is high. The method comprises following realization steps: (1) constructing a directed graph of the network; (2) generating an adjacent matrix corresponding to the directed graph; (3) quantitating influence factors of user's nodes on information dissemination; (4) quantitating topic relevancy of user's nodes; (5) quantitating information dissemination capacity of user's nodes; and (6) recognizing key nodes. The recognition method of key nodes of a social network in a topic field has following beneficial effects: the method depends on a local topological structure and introduces influence factors of user's nodes on information dissemination and the concept of topic relevancy of user's nodes in the real social network so that computing complexity is reduced and key nodes of the social network in the topic field are effectively recognized.

The invention provides a recognition method for new words, which comprises the following steps: segmenting a scientific and technical text into sentences, and establishing a mechanized dictionary; matching and segmenting the short sentences segmented from the text in vocabularies, then, atomically segmenting the remaining sentence strings, and automatically extracting the two-character words, three-character words and multi-character words; and sequencing the extracted words with the statistical method, and evaluating the sequenced words to obtain new words. The method can greatly increase the recognition accuracy and the comprehensiveness of the new words.

The invention provides an end-to-end tendency scene character detection and recognition method and an end-to-end tendency scene character detection and recognition system. The method comprises the steps: collecting pictures marked with object types, constructing a training data set and a test data set, and carrying out the preprocessing of the data sets for training a feature extraction network; collecting pictures marked with character positions, constructing a training data set and a test data set, and preprocessing the data sets to train a network of a character detection part; pictures marked with character positions and character contents at the same time are collected, a training data set and a test data set are constructed, and corresponding picture preprocessing is carried out to train a neural network of a character recognition part; inputting the picture into a convolutional neural network; extracting a shared convolution feature map through a feature extraction network; andmapping predicted character position coordinates to the shared convolution feature map by the character detection part maps by using the shared convolution feature map, cutting off a feature map blockcorresponding to the character part in the image and converting the feature map block into a feature sequence; finally decoding the feature sequence into a readable character sequence. Since the convolution characteristic spectrum is only calculated once, the intermediate redundancy process is avoided. The speed of the whole scene character detection and recognition system is improved.

The invention relates to the technical field of mode recognition, and in particular to an attribute learning and Riemannian manifold combined pedestrian re-recognition method. According to the method,a model with stronger generalization ability is disclosed through researching representations and semantic attributes of pedestrians; and through importing pedestrian attribute labels, using a convolutional neural network model and using multi-target loss functions, the model not only needs to predict pedestrian IDs, but also needs to predict various correct pedestrian attributes. Pedestrian pictures are converted into output tensors of a convolution layer through the trained convolutional neural network, so as to respectively calculate a covariance descriptor of each pedestrian picture. Thecovariance descriptors are utilized to carry out feature fusion so as to eliminate feature redundancy, so that measurements of Riemannian manifolds where the descriptors are located are researched andthen more correct similarity calculation is realized.

For carrying out structural sorting on the online education courses and test resources, marking is carried out through the knowledge point-based tag system, and then based on the location and the relationship of the knowledge points in the knowledge map, structural description of the courses and test resources is formed. The behavior data of a student in browsing, retrieving and learning online courses as well as conducting various online tests is collected, filtering, cleaning, sorting, associating and recalculation are carried out according to a certain method, and an effective behavior description log of the individual student learning process is established. Based on the structural description of resources and the behavior log description of the individual student, behavioral marking, performance quantification and ability rating are carried out on the student from the perspective of knowledge points, and a preliminary ability quantification model is established.

The embodiment of the invention discloses a diatom detection and recognition method based on a deep learning algorithm. The diatom detection and recognition method is used for solving the problems oflow recognition efficiency and inaccurate recognition caused by too many types and complex backgrounds in the diatom inspection process. The embodiment of the invention comprises the following steps:S1, obtaining various diatom types of images, and making a data set according to a Pascal VOC2007 data set format; S2, training a target detection model for various diatom targets through a deep learning target detection algorithm; S3, using the trained Faster R-CNN network model to detect diatom targets in an image to be detected, and the image entering the convolutional layer of the Fast R-CNN network model, inputting the feature map output by the last shared convolutional layer into an RPN network model to generate candidate regions where targets may exist, outputting the central coordinates and the width and the height of the regions, and inputting the features of the candidate regions into subsequent classification and frame regression part in the Fast R-CNN, to obtain a target type and refined position information.

The invention discloses a GRU (Gated Recurrent Unit)-CRF (Conditional Random Fields) conference name recognition method based on a language model. The GRU-CRF conference name recognition method basedon the language model comprises two parts: one part involves a GRU-based language model, and the other part involves a GRU-CRF-based recognition model. An end-to-end recognition model which does not require feature engineering and domain knowledge is obtained by using a labeled supervision data training and labeling model GRU-CRF. Unsupervised training is performed on the LM (Language Model) by using a large amount of unlabeled data, and a word vector is acquired from the LM obtained by unsupervised training as an input of GRU-CRF, so that the supervised training effect can be improved, and the generalization ability of the recognition model can be improved, therefore, that a named entity recognition model with a relatively good effect is trained from a small amount of labeled corpus becomes possible. Experimental results show that an LM-GRU-CRF method has the best effect on a self-built corpus, and for other named entity recognition tasks that lack the labeled corpus, the method can be used for improving the effect of the model.

Provided is a pose recognition method of a leaner. The method comprises steps of separation of a portrait and a background; using mathematical morphological operation to extract a profile image of a learner for a binary image; using a Zernike matrix to carry out feature extraction; adopting a support vector machine to train the feature vectors; and recognizing poses of the learner. According to the invention, by introducing the Levy flight mechanism in a bee colony algorithm, according to different flight step lengths, different searching modes are adopted, so the variety of the population can be enriched; premature convergence and falling into the local optimum can be avoided; a dynamic weighting mechanism is adopted in the bee colony algorithm, and an individual bee searching mode is adjusted according to the evolutionary rate of the population, so global searching and local searching ability are dynamically balanced; and the method is advantaged by good segmentation effects and high recognition rate, and can be used for recognition of poses of the learners and recognition and classification of other images.

In certain embodiments, training of a prediction model (e.g., recognition or other prediction model) may be facilitated via a training set generated based on one or more logos or other graphics. In some embodiments, graphics information associated with a logo or graphic (e.g., to be recognized via a recognition model) may be obtained. Training media items (e.g., images, videos, etc.) may be generated based on the graphics information, where each of the training media items includes (i) content other than the logo and (ii) a given representation of the logo integrated with the other content. The training media items may be processed via the recognition model to generate predictions (related to recognition of the logo or graphic for the training media items). The recognition model may be updated based on (i) the generated predictions and (ii) corresponding reference indications (related to recognition of the logo for the training media items).

An infant cry recognition method includes: collecting voice data, and intercepting a voice data segment of preset time length; calculating two or more audio feature vectors included in the voice datasegment; recognizing the audio feature vectors according to a preset recognition model, and sending a recognition result to a monitoring terminal. The recognition result is acquired by recognizing according to the audio feature vectors, so that the recognition result is more accurate and reliable, and improving of infant cry recognition accuracy is facilitated.

The invention provides a method and a device for recognizing problem types and a method and a device for establishing recognition models. During a process of establishing a question recognition model, whether a text fragment belongs to feature fragments of a class is judged on the basis of the absolute frequency of occurrence and the relative frequency of occurrence compared with other classes of the text fragment containing interrogatives or requirement words in the class, and accordingly the type to which the text fragment containing the interrogatives or the requirement words belongs can be positioned accurately, and the accuracy of recognition of question types can be improved. In addition, the question type recognition model can further be used for reclassifying training corpuses to serve as updated training corpuses, and the question type recognition model can be optimized gradually through this kind of iteration mode. If question type recognition and searching are performed on the basis of the mode, the searching accuracy can be further improved.

The invention discloses a human motion intention recognition method based on an intelligent lower artificial limb. The method solves the problem that a traditional human motion intention recognition method is hysteretic. Before the motion mode of the affected side wearing the artificial limb is switched, the motion intention of a handicapped patient is accurately recognized according to time sequence data of a sensor which is embedded into the artificial limb or bound to the normal side. Besides, compared with a traditional intention recognition method where various sensors are used, only onesensor, namely an inertial sensor, is used in the human motion intention recognition method to accurately recognize the human motion intention. The method includes the following steps that intention recognition data is acquired, and a database is built; the intention recognition data is preprocessed; the intention recognition data is subjected to feature extraction; classification model selection,model training and intention recognition are completed. An experimental data recognition method is provided for the motion intention recognition method of the intelligent lower artificial limb, so that the development of the intelligent artificial limb is promoted, and lower extremity amputation patients can be better served.

The invention discloses a point-of-interest (POI) data error type positioning method. According to the method, word segmentation processing is performed on a basic field of a POI data pair according to set layering parameters of the basic field of a POI, and the number of layers where all segmented words forming the basic field are located is acquired; the POI data pair comprises POI data input by a user and POI original data corresponding to the POI data input by the user; the basic field comprises at least one subfield; a similarity value between the subfields of the POI data pair is calculated according to the number of layers where all segmented words forming the subfields are located; and when it is judged that the similarity value between the subfields of the POI data pair belongs to error threshold values according to the set error threshold values corresponding to different error types of the subfields, the error type of the subfields of the POI original data is positioned. Meanwhile, the invention furthermore discloses a POI data error type positioning device and a POI data repeated recognition method and device.

The invention provides a target character distance measurement learning-oriented character re-recognition method, and relates to the technical field of video processing. According to the method, a non-target query character is imported onto a target task data set to serve as an auxiliary constraint, and under a distance measurement and transfer learning framework, distance measurement orienting tothe target query character is online learnt, so that the character re-recognition problem of comparing multiple frames of images is solved. According to the method, the influences caused to statistical model training and learning on target task data sets by diversity and difference of data sets when third-party character re-recognition task data sets are imported are avoided, the problem of statistical model overfitting caused by less training data is relieved, loss functions are defined at near distances of target character feature instances, the problem that an in-class variance of a same character feature instance is greater than an inter-class variance is overcome, and character re-recognition tasks for comparing multiple frames of images are solved.

The invention discloses a character detection and recognition method for Chinese historical literature dense texts, which comprises the following steps of: (1) making data acquisition: acquiring historical literature images, and manually labeling the historical literature images; (2) pre-processing data: performing column segmentation on the vertical projection of the historical literature images,and cutting the vertical text in the historical literature according to columns; (3) constructing and pre-training a convolutional neural network for single-row text recognition; (4) constructing a convolutional neural network for performing character detection on the single-row text, sharing shallow parameters with the convolutional neural network for performing single-row text recognition, andperforming training at the same time; and enabling the text detection convolutional neural network to identify the text information provided by the convolutional neural network by using the text, andfinely adjusting the detection position, so that the single text position of the dense text in the historical literature can be accurately detected. According to the invention, the convolutional neural network is adopted to realize text recognition, the guidance information of the text recognition classifier is fully utilized, and the detection effect can be more accurate.

The invention provides a flammulina velutipes head sorting recognition method based on LBP features and deep learning, and the method comprises the steps of 1, collecting the flammulina velutipes headpictures, and dividing all the flammulina velutipes head pictures into a training set and a test set; 2, changing the flammulina velutipes head pictures in the training set, and storing the flammulina velutipes head pictures before change and the changed flammulina velutipes head pictures in the training set as training data; 3, extracting an LBP feature a of the training data; 4, extracting a depth feature b in the training data by using the convolutional neural network; 5, fusing the LBP feature a and the depth feature b after dimension reduction to obtain a fusion feature c; 6, inputting the fusion feature c into a classifier for classification to obtain a trained model; and 7, inputting the flammulina velutipes head picture in the test set into the trained model to obtain a predictedvalue, and comparing the predicted value with a real value to calculate the accuracy. According to the invention, the accuracy and the efficiency of the flammulina velutipes head classification are improved.

The invention discloses a data stream exchanging and multiplexing system and a data stream exchanging and multiplexing method suitable for multi-stream regular expression matching. The system comprises a priority adding module, a programmable storage module, an information exchanging module, an exchange scheduling module and a single-stream REM (Recognition Memory) module, wherein the priority adding module is used for judging the characteristics of input data streams and endowing corresponding data stream priorities and data stream serial number information; the programmable storage module is connected with the information exchanging module, and is used for storing the data streams, the corresponding data stream priorities, the corresponding data stream serial numbers and data stream waiting time information; the information exchanging module is connected with the priority adding module, the programmable storage module and the exchange scheduling module respectively; the exchange scheduling module is connected with the information exchanging information, and is used for dynamically selecting data streams to be processed and adjusting the exchanging lengths of the data streams; and the single-stream REM module is connected with the information exchanging module, and is used for performing regular expression matching on data streams transmitted by using the information exchanging module. According to the system and the method, simultaneous processing of multiple data streams can be supported, and high flexibility and a high resource utilization ratio are achieved.

The invention provides an inscription rubbing-based calligraphy character segmentation and recognition method. The method comprises the following steps of 1, processing images in a training set; 2, pre-processing a training sample; 3, reading an image processed in the step 2 for model training, wherein a feature matrix MLH and a feature vector matrix Me are models trained according to rubbings; 4, conducting the word segmentation for the entire image, and segmenting calligraphy characters out of the image; 5, judging each character segmented in the step 4 belongs to which character in rubbings, obtaining a dimensionality reduction feature matrix MP of the image, calculating the similarity, calculating the distance Dist between MP and each row of MLH, and determining that characters in one row of MLH that is smallest in distance to be to-be-recognized characters. According to the technical scheme of the invention, the inscription rubbing-based calligraphy character segmentation and recognition method can effectively realize the segmentation of calligraphy characters, and is high in recognition accuracy and fast in recognition speed.

The invention relates to a pseudo differential heart beat and abnormal heart beat recognition algorithm based on misclassification and supervised learning. The algorithm comprises the following stepsthat 1, electrocardiogram data of an electrocardiogram signal database labeled with an existing heart beat type is adopted for recognizing R peaks and extracting heart beat features; 2, by comparing with the R peaks labeled in the database, the misrecognized R peaks and noise heart beats labeled in the database are classified to be pseudo differential heart beats; 3, heart beat features of eight types of the heart beats of the pseudo differential heart beats, normal heart beats, ventricular premature beats, ventricular escape beats, supraventricular premature beats, supraventricular escape beats, ventricular fusion beats and pacemaker heart beats in the database are extracted to serve as training data; 4, a supervised learning method is used for training the training data to be an eight classification model; 5, test data in real-time dynamic electrocardiogram data is extracted and input into the classification model to obtain the heart beat classification result. The algorithm is suitable for recognition of dynamic electrocardiogram long time electrocardiogram data pseudo differential heart beats and other multiple types of abnormal heart beats.