33results about How to "Improve recognition performance" patented technology

a method for assessing memory in a subject include the steps of presenting to the subject a list of items to be retrieved from memory by the subject, having the subject recognize the presented items from memory, determining the subject's response speed to each of the recognized repeated items and analyzing a plurality of the response speeds for the recognized repeated items. The items which presented to the subject are intermixed with repetitions of the items being tested for recognition. The subject is tested to determine if he recognizes each repeated item as being a repeated item. The response speed for each of the recognized repeated item is the time required between when the subject is shown a repeated item and when the subject responds that he recognizes the repeated item.

The invention provides a mode training method based on ensemble learning and a mode indentifying method. The mode training method comprises the following steps of: 1) carrying out dictionary learning on training samples to generate a redundant dictionary; 2) utilizing the redundant dictionary to carry out sparse encoding on the training samples to obtain a sparse encoding coefficient of each training sample; 3) carrying out sparse subspace division on all the training samples according to the sparse encoding coefficients; and 4) carrying out sub-model training on the training sample in each sparse subspace to obtain a sub-model for classifying. According to the mode training method based on the ensemble learning and the mode indentifying method, provided by the invention, higher indentifying performance can be obtained; and meanwhile, the training efficiency and the detection efficiency can be obviously improved.

The invention relates to a micro expression recognition method based on multi-feature multi-task dictionary sparse migration learning. The micro expression recognition method comprises a training phase and a test phase. According to the micro expression recognition method provided by the invention, a macro expression and a micro expression are projected into a public space by means of projection,and in order to simplify the calculation and to improve the efficiency, sparse dictionary representation is performed on the projected data; in order to further reduce the data difference between thetwo domains, dictionaries of the two domains are reconstructed from each other to realize the relevance of the dictionaries, so that projected sparse representation matrixes generate greater correlation; in order to fully express the characteristics of the micro expression, in the micro expression recognition method provided by the invention, four kinds of different features are extracted for themicro expression, and the optimal combination is achieved by the selection of multiple features; and in order to highlight the detailed expression of the micro expression, the multi-task idea is imported by the micro expression recognition method provided by the invention to further enhance the recognition effect.

The invention discloses a network packet protocol identification method. The method includes a protocol configuration step and a packet identification step. The protocol configuration step includes: storing characteristic information of protocols; establishing a protocol tree; and establishing a table of characteristic values and judgment logic. The packet identification step includes: acquiring a data packet to be recognized; selecting the protocol tree for identifying the data packet protocol; and comparing values of keywords, read from the data packet, to the table of characteristic values and the judgment logic so as to identify the data packet protocol. The invention further provides a network packet protocol identification system. The table of characteristic values and the judgment logic are established according to characteristic information of all protocols, the protocol used by the data packet can be quickly found by single table look-up, and accordingly the method and system are high in identification efficiency. Protocol identification information of a new protocol needs to be added only when the new protocol is added, so that the method and system are highly extensible.

The invention discloses a court judgment document-oriented multi-scale learning character recognition method and system. The method comprises the steps of: acquiring a to-be-recognized court judgmentdocument image, and extracting a seal area; constructing a seal generation network model based on an adversarial network by taking the maximum difference of the target RGB values as an objective function and taking the cyclic consistency loss as a constraint condition, performing seal trace RGB value conversion on the seal area by adopting the trained seal generation network model, and deleting the seal trace of the converted seal area; and carrying out feature extraction on the court judgment document image with the seal trace deleted, respectively carrying out global target detection and local detail detection on an obtained feature map, combining the obtained masks and progressive masks of text candidate boxes, then training a constructed text detection model, and obtaining a text recognition result by using the trained text detection model. The problems of seal trace shielding in the text image and detection of super-long and super-short texts are effectively solved.

The invention discloses a motor vehicle two-dimensional identification management system and application method thereof. The system comprises hardware and identification management program software; wherein the hardware includes a two-dimensional code laser marking machine, a two-dimensional code scanner and a computer server, the computer server transmits printing information to the two-dimensional code laser marking machine, the two-dimension code scanner transmits scanned data to the computer server, the identification management program software is installed in the computer server and is characterized by comprising a database module, a two-dimensional code generating module, an information processing module and a two-dimensional analysis module. As identification mode of two-dimensional code is adopted, not only information of license plate number of motor vehicle is contained, but also identification information of engine number, rack number and the like and dynamic management information are contained, thus greatly improving identification property; and light scribe printing mode is adopted for glass two-dimensional code graphical information of vehicles, thus effectively avoiding dropping in the course of time and man-made wiping and replacing.

The invention discloses an SAR target small sample identification method based on a graph attention network. The method mainly solves the problem of poor recognition rate under the condition of lack of training data in the prior art, and adopts the scheme of selecting SAR images containing a radar target, suppressing speckle noise of the SAR images, and dividing the SAR images after noise reduction into data with a label and data without a label; training an auto-encoder by using the denoised image to obtain feature vectors of all SAR images; obtaining an initial adjacency matrix by utilizingvector similarity on the premise of a small amount of label data; and setting a graph attention network, iteratively training the graph attention network by utilizing all the feature vectors until anerror function of the network is converged, and outputting a finally predicted node label matrix to realize label-free data identification. According to the method, a small number of known types of SAR targets can be utilized to predict the types of a large number of other unknown targets, the prediction accuracy is high, and the method can be used for radar target recognition under the conditionof small samples.

The invention discloses an illumination processing method for human face images, which belongs to the field of image processing. The method comprises the following steps of: 1, inputting the human face images and normalizing the human face images to achieve a fixed size; 2, performing illumination processing on the normalized human face images by different illumination processing methods respectively; and 3, fusing the images which are subjected to the illumination processing and outputting a fused image. The human face images are processed by the single-dimensional Retinex illumination processing method of a horizontal edge filtering kernel, the horizontal edges of the human face images are highlighted, the images acquired by the different illumination processing methods are fused according to the weight coefficients of the images, and the fused images keep the advantages of images before fusion, so that the performance of human face identification is improved, and the identification rate of a human face is increased.

The invention discloses a self-adapting method of a DNN acoustic model based on personal identity characteristics. The method solves the problems of easy over-fitting, poor personal identity representation capability and low robustness in the self-adapting training. The method comprises the following steps of: extracting personal identity characteristics, and inputting the MFCC characteristics asthe DNN model of a non-specific speaker; establishing a GMM-HMM speech recognition system; building a DNN-HMM baseline system of the DNN acoustic model with a plurality of hidden layers; and carryingout self-adapting training on the individual identity characteristics of the DNN acoustic model layer by layer to obtain the DNN acoustic model which has self-adapting capability to a specific speaker. In personal identity characteristic extraction, a weight matrix decomposition of the last hidden layer of DNN model is replaced by VAD technology. According to the method, a small amount of speakerdata is fully utilized to adjust the model parameters so that the recognition accuracy rate of the specific speaker is improved. And the complexity is low, and the recognition performance is obviouslyimproved. The method is used for intelligent systems related to speech recognition or communications, medical treatment, vehicle mounting and the like.

Speech recognition performance is improved without changing a speech recognition engine. A speech data generation section generates, from speech data for which speech recognition is to be performed, a plurality of pieces of speech data whose starting positions of the non-speech regions differ. A speech recognition engine performs speech recognition by using each of the pieces of speech data. A totaling / comparison section provides the most numerous recognized result from among a plurality of obtained recognized results.

The invention provides a Chinese named entity recognition model, a creation method thereof and a method applied to the field of network space security. The application of the Chinese named entity recognition model is based on transfer learning and a deep neural network, and comprises the following steps: firstly, training a Bert-BiLSTM-CRF model on four general data sets recognized in the field ofChinese named entity recognition, and fully learning general knowledge features; performing model migration, training the TBBC (Trans-Bert-BiLSTM-CRF) model subjected to transfer learning on a self-labeled network space security domain data set, and obtaining a trained TBBC model; learning to obtain features of the domain knowledge, outputting the model, finally obtaining a TBBC model with a practical application value, and performing Chinese named entity identification. Tests show that the accuracy, recall rate and F1 value of the obtained TBBC model are obviously improved; the Chinese namedentity recognition performance is greatly improved, and the practical problems of insufficient training data and low recognition performance when named entity recognition tasks are carried out in a specific field can be effectively relieved.

The invention belongs to the technical field of radar image processing, and particularly relates to a small sample SAR image target recognition method based on an improved prototype network, which can be used for SAR automatic target recognition under a small sample condition. The method comprises steps of acquiring a training sample set and a to-be-recognized small sample SAR image set, constructing a deep convolution-bidirectional long-short time prototype neural network, namely an improved prototype network, adopting a training support set and a training query set to train the deep convolution-bidirectional long-short time prototype neural network, and obtaining a target recognition result of the small sample SAR image. According to the method, a new network structure is constructed, so that the problem that each type of to-be-recognized target needs hundreds of or even more training samples in a traditional recognition method is solved, and the targets are effectively classified and recognized by using a small number of labeled samples in each type.

The invention provides a cross-angle gait recognition method bases on multi-coupling discrimination local block alignment, including online training and offline testing. Compared with other recognition methods, the method provided by the invention is better in recognition performance, and does not require completely-controlled environments and complicated camera calibration. In addition, the method provided by the invention does not require inverse operation of a matrix, so that the operand is reduced, and the small sample problem is reduced. The method is a cross-domain gait recognition method with high robustness properties.

The invention discloses a method for generating pedestrian re-identification data under different illumination conditions based on an adversarial network. The method comprises the following steps: firstly, acquiring videos of different time periods in a monitoring video, namely monitoring videos under different illumination conditions; Extracting representative monitoring images from the acquiredmonitoring videos, and marking the representative monitoring images into three different categories of'too dark, normal and too bright 'according to illumination conditions; Sending the images with the illumination condition labeling information into a generative adversarial network for training, wherein the trained generative network is used for generating data under different illumination conditions; And generating the pedestrian re-identification data into data under different illumination conditions through the trained generation network; And finally, adding the obtained data as a data augmentation form into pedestrian re-identification model training, and at the moment, enabling the model to identify the characteristics of the same target under different illumination conditions, i.e.,enhancing the identification performance of the model on the target under different illumination conditions.

The invention provides a method for extracting characteristic of a handwritten Chinese character image. The global handwritten Chinese character image is used as a characteristic extraction area; furthermore, a Chinese character image area is divided by an elastic network; scale invariability characteristic conversion method is employed to carry out dynamic statistics of gradient direction information of relative areas on each network, thus gaining the characteristic of a handwritten Chinese character. The method randomly selects 500 samples from a HCL2000 handwritten Chinese character sample database to carry out a training and selects 200 non-repeated samples to carry out a recognition test; the recognition result when the method is used for gaining characteristic is that the hit ratio of a firstly selected character is 96.061% and the hit ratio of the first 10 candidate characters is 99.688%.

The invention puts forward a cervical cancer focus analysis method based on cell image recognition. The cervical cancer focus analysis method of the invention has the beneficial effects that: pixel grade features with abundant feature expression and super pixel grade features with certain semantic features are combined to obtain level grade features, then the level grade features are used for training a random forest to perform segmentation of cells, thereby well describing a microstructure with intricate and complex parts, and improving a recognition performance; in addition, automatic analysis is performed on an image from a cell level grade, a cervical cancer focus can be captured accurately, thus targeted treatment can be performed, and a high automation degree is achieved.

The invention discloses a network community user identification method and device, and relates to the technical field of machine learning and user identification. The method comprises the following steps: extracting a first N-gram feature of the pre-collected network community text data and generating a first word vector; Taking the cross entropy as a cost function, and training the first word vector by adopting a deep neural network to obtain a text content analysis model; Analyzing the text data of the to-be-identified user by using a text content analysis model to obtain a content attribute; Analyzing the behavior data of the to-be-identified user by using a preset behavior analysis model to obtain behavior attributes; And identifying the to-be-identified user according to the content attribute and the behavior attribute. In the invention, N-gram features are introduced, the target user identification is carried out based on the the user content and user behavior two-dimensional data, and identification performance and accuracy are greatly improved compared with existing single-dimensional data user identification.

The invention discloses a layer-by-layer channel selection method for voice recognition of a self-organizing microphone, and the method is based on a conformer voice recognition framework, and the method comprises: (1), an encoder-decoder framework is adopted, an encoder is based on a Conformer framework, a decoder is based on a Transformer framework, and a multi-head attention mechanism is introduced into an encoder-decoder module; (2) for a single-channel voice recognition system, clean voice is adopted for independent training; and (3) for a multi-channel voice recognition system, the voice of each channel is encoded and then the same decoder is shared, a multi-layer flow attention mechanism is trained, and the channels are screened layer by layer. Under a large-scale self-organizing microphone array, compared with other flow attention-based methods, the method provided by the invention is higher in speech recognition accuracy and lower in calculation complexity.

The invention relates to an audio signal processing method and device and a storage medium. The method comprises: acquiring, by at least two microphones, audio signals sent by at least two sound sources respectively so as to acquire original noisy signals of the at least two microphones respectively; for each frame in the time domain, obtaining respective frequency domain estimation signals of atleast two sound sources according to the respective original noisy signals of the at least two microphones; dividing a predetermined frequency band range into a plurality of harmonic subsets, each harmonic subset containing a plurality of frequency point data; determining a weighting coefficient of each frequency point contained in each harmonic subset according to the frequency domain estimationsignal of each frequency point in each harmonic subset; determining a separation matrix of each frequency point according to the weighting coefficient; and based on the separation matrix and the original noisy signals, obtaining audio signals emitted by the at least two sound sources respectively. Through the method provided by the embodiment of the invention, the voice quality of the audio signals can be improved.

The invention provides an attack program identification method based on a vulnerability attack database and a decision tree. The method comprises the following steps: step 1, constructing a vulnerability attack database facing network traffic; step 2, preprocessing a to-be-detected attack program data set to obtain a data set with higher quality and stronger representativeness; step 3, selecting key feature subsets in the vulnerability attack database and the attack program data set to be detected through a secondary feature selection method based on principal component analysis; step 4, training the key feature subsets selected in the vulnerability attack database by using a fast decision tree algorithm to construct a classification model; and step 5, completing the identification of attack programs according to the results of the step 3 and the step 4.

The invention provides a voice noise filtering method and device, electronic equipment and a medium, and relates to the technical field of voice recognition. The voice noise filtering method comprises the following steps: carrying out segmentation operation on input voice information to obtain segmented voice; performing hierarchical clustering operation based on the human voice features in the segmented voice to obtain an initial cluster; correcting the initial clustering cluster based on K-means clustering, and determining a corrected clustering center; according to the distance between the segmented voice and the clustering center and the distance distribution, identifying noise in the segmented voice; and performing a noise filtering operation on the recognized noise to obtain the filtered voice information. Through the technical scheme disclosed by the invention, the filtering effect on invalid audios such as polyphonic ringtones, laughter, cough sound, background human voice and channel noise can be improved, so that the recognition performance of a voice recognition system provided with the noise filtering module disclosed by the invention is improved, and the use experience of a user is improved.

The embodiment of the invention provides a method and device for generating a texture image and a composite image executed by a computer, and the method comprises the steps: applying a basic graphic function to a coordinate value of any first position in a to-be-generated image, and obtaining a basic value; for the first position, generating a disturbance noise value at least according to the noise disturbance function; applying an adjustment function to the sum of the basic value and the disturbance noise value to obtain a pixel value of the first position; and generating a texture image according to the pixel value of each position. By using the method, a large number of complex, non-repeated and low-cost texture images are generated, and the method can be used for training an image recognition model. A composite image can be generated by combining the texture image and the text image, and position information of each text in the composite image is recorded as text position label data of the composite image. And the synthesized image is used as a verification code to train a verification code recognition model, so the recognition capability of the recognition model is effectivelyimproved.

The invention discloses an end-to-end bone and air conduction voice combined recognition method, which comprises the following steps of: firstly, acquiring synchronous air conduction and bone conduction voice data to construct a data set, and outputting a corresponding text; performing data enhancement and acoustic feature extraction on the air conduction and bone conduction voice signals; then, a Conformer-based end-to-end deep neural network model is built, and the Conformer-based end-to-end deep neural network model is composed of three parts, namely two branch networks for processing air conduction voice and bone conduction voice, and a fusion network based on multi-mode Transducer; and then training the neural network, and finally obtaining a corresponding recognition result through the trained network. Compared with a traditional method that voice recognition is carried out only through air conduction voice signals, the combined recognition method can remarkably reduce the error rate of voice recognition, and the overall recognition performance of the system is improved.

The invention discloses a static gesture intention recognition method and system based on dynamic feature assistance and a vehicle. The method comprises the following steps: step 1, hand detection; the method comprises the following steps of: 1, extracting dynamic features, 2, hand tracking, 3, static gesture classification, 4, spatial position judgment, 5, continuous frame recognition, 6, dynamic feature extraction and judgment, 7, static gesture judgment output and 8, in-vehicle application event response. According to the method, the static gesture actions are classified based on the hand action information of the user, the static gesture intention of the user can be recognized more accurately by combining the forward hand action trend characteristics in the static gesture control process, the misrecognition situation caused by the fact that similar gestures are static and free of operation or unconscious random waving is reduced, and the user experience is improved. And the false triggering rate is reduced, so that the cockpit gesture control function experience of the user is improved.

The invention relates to a cross-domain pedestrian re-identification method based on median clustering and global classification. The method includes the following steps: 1, processing source domain data by using a style migration model, and then performing supervised learning on a source domain by using a feature extraction model F, namely pedestrian identity classification; 2, performing feature extraction on the target domain data by using a feature extraction model, and then clustering the features by using median stable clustering to obtain a clustering result and a corresponding pseudo tag; and 3, respectively optimizing a global feature classifier and a feature extraction network by utilizing a clustering result, wherein the feature extraction network takes the classification precision of the classifier as an optimization target. The global feature classifier and the feature extraction model designed by the invention are subjected to collaborative optimization after clustering, and finally distance distribution of positive and negative sample pairs is separated in a global range, so that the recognition performance of the model is improved.

The invention discloses a named entity identification method and a device thereof, a computer readable storage medium and a processor. The method comprises the steps of obtaining text data; performing feature mapping on the text data to obtain a first text vector and a second text vector; and respectively inputting the first text vector and the second text vector into two channels of an identification model for named entity identification to obtain an identification result of the text data. According to the method and the device, the technical problem of relatively low identification performance caused by word embedding deviation easily occurring in a named entity identification process in related technologies is solved.

The invention discloses a negative emotion detection voice emotion recognition method based on context amendment, and belongs to the field of voice signal processing. The negative emotion detection voice emotion recognition method includes the steps of firstly collecting emotion data, conducting feature extraction on each voice, using the principal component analyzing dimensionality reduction technology to reduce feature dimensionalities, adopting two kinds of classifiers based on a Gaussian mixture model to judge four kinds of emotions respectively to obtain an emotion vector at the current moment, finally amending the emotion vector at the current moment according to an emotion vector at a previous moment and perceiving task performance at present, and accordingly obtaining final voice emotion recognition results. According to the negative emotion detection voice emotion recognition method, the recognition performance of a voice emotion recognition method singly with a Gaussian mixture classifier can be effectively improved, and a negative emotion state related to the recognition process is effectively detected. Particularly in the special work environments such as spaceflight and navigation, the negative emotion detection voice emotion recognition method has important application value on detecting and adjusting negative emotions of workers.

The invention belongs to the technical field of artificial intelligence, and provides an image re-identification method and device, electronic equipment and a storage medium. The image re-identification method comprises the steps of: obtaining a to-be-identified image; based on a preset feature extraction model, extracting global re-identification features and local re-identification features of the to-be-identified image; performing fusion processing on the global re-identification features and the local re-identification features of the to-be-identified image to obtain fusion features; and carrying out classification processing on the fusion features to generate re-identification features, used for image re-identification, of the to-be-identified image. According to the image re-identification method and device, the electronic equipment and the storage medium, the features of multiple scales can be extracted from the to-be-identified image based on the preset feature extraction model to perform image re-identification, the robustness of image re-identification is enhanced, and the identification performance of image re-identification is improved.

The invention discloses a device control method, device, storage medium and device for children's accent recognition. The method obtains the user's voice command, and according to the voice command, recognizes the interaction mode through a wake-up word recognition model based on deep recognition technology, and can accurately Distinguish whether the speaker is a child; when the current interaction mode is in the child mode, the voice command of the user will be recognized, and at least two text information corresponding to the voice command will be obtained, and the text information will be spliced. The splicing results are processed in natural language to obtain interactive information, and perform device operations corresponding to the interactive information. In this embodiment, when the current interaction mode is in the children's mode, the text information corresponding to the voice command is no longer regarded as independent information, but after splicing it, natural language processing is performed on the splicing result, thereby improving the child's voice. The recognition performance improves the accuracy of performing corresponding equipment operations.

The invention provides an image processing system. The image processing system comprises a low-light CMOS image sensor, an infrared focal plane detector and a processor. The low-light CMOS image sensor and the infrared focal plane detector are respectively connected with the processor; the low-light CMOS image sensor is used for acquiring a visible light image and sending the visible light image to the processor; the infrared focal plane detector is used for acquiring an infrared image and sending the infrared image to the processor; and the processor is used for obtaining a color fusion imageaccording to the visible light image and the infrared image. The visible light image and the infrared image are processed by adopting a fusion algorithm, multi-source information is comprehensively utilized, scene understanding is enhanced, a target is highlighted, an image with the performance better than that of any source image can be obtained, and the target can be more quickly and accuratelydetected under the hidden, camouflaged and confusing military backgrounds. And the fused image is displayed in a natural form suitable for human eye observation, so that the human eye recognition performance can be obviously improved, and the fatigue feeling of an operator is reduced.