32results about How to "Good classification recognition effect" patented technology

The invention relates to a twitter identification method based on self-adaption energy detection under a complex environment. The method is characterized by comprising the following steps of firstly, dividing twitter into three frequency bands according to the frequency distribution condition of the twitter; secondly, detecting twitter acoustical signals containing noise through a self-adaption energy detection method based on noise power spectrum estimation and the foreground sound existence probability, and screening out the useful twitter signals; thirdly, extracting Mel-scale WPSCC features from the useful twitter signals according to distribution of the Mel scales; fourthly, conducting modeling, classification and identification on the extracted Mel-scale WPSCC features and extracted Mel-scale MFCC features through an SVM classifier. The WPSCC features extracted through the method have a better anti-noise function and are better in identification performance after undergoing self-adaption energy detection, the method is more suitable for twitter identification under the complex environment and has a good classification and identification effect on the twitter sound containing noise under the complex environment.

The invention discloses a facial expression recognition method based on random forests. The facial expression recognition method based on the random forests comprises the step of extraction of a displacement feature of an AAM, the step of extraction of AUs in a facial expression sequence, the step of training of a facial expression classification model and the step of facial expression recognition. According to the facial expression recognition method, the novel AAM displacement feature is provided to be used for training and learning the AUs, and finally facial expression recognition is carried out by depending on the AUs. Compared with other feature representations in identification of the same classification, the facial expression recognition method based on the random forests better describes expression information and changing process information contained in the expression sequence. The random forests are used for facial expression recognition for the first time, and the random forests in the method have a better classified recognition effect in the field compared with a frequently used support vector machine (SVM) method at present. For the aspect of CK and AU recognition of databases, the facial expression recognition method based on the random forests can achieve a perfect recognition effect.

The invention relates to an industrial-grade intelligent surface defect detection method, which comprises the following steps: constructing and training to generate a twin generative adversarial network GAN, repairing an input image into a normal sample through an improved GAN network, and comparing the output with a manually labeled positive sample through a twin CNN network to obtain a difference, which is a defect. The twin generative adversarial network designed by the invention does not need a large number of samples and does not need to perform data amplification, can solve the problem of small sample size of common industrial products, and reduces the occurrence of an overfitting phenomenon caused by the small number of samples and zero samples in deep learning, so that the defect detection during the development of products and new products with small defect sample size becomes possible. A cross alignment loss function CA and a distribution alignment loss function DA are used to enhance the relationship between two network outputs, and a good classification and identification effect is obtained. The model training speed is improved through an Attention mechanism and a hardware GPU, so that industrial rapid deployment becomes possible.

The invention discloses a crop classification and identification method under a strong noise background. The method comprises the following steps of: shooting a plurality of pictures of various cropsby using a multispectral camera to form a picture set; obtaining the NDVI value of each pixel point and segmenting the NDVI value into plant areas; replacing the non-plant area with a pure color background to highlight the plant area, performing image preprocessing to form a multispectral data set, and dividing the multispectral data set into three data sets, namely training, testing and verifying; inputting the training data set into a preset convolutional neural network model for training through a transfer learning method to obtain a convolutional prediction neural network model, and inputting the test data set into the convolutional prediction neural network model for accuracy testing to obtain a qualified convolutional prediction neural network model; and inputting the verification data set into the convolutional prediction neural network model, carrying out classification and identification on crops in the verification data set, and obtaining a classification result. According tothe method, the influence of a strong noise background on crop classification and recognition is reduced, and the recognition efficiency and prediction capability of the model are improved.

The invention discloses an extension neural network pattern recognition method based on priori knowledge. The method includes the following steps that (1) a training sample set and a knowledge base are prepared; (2) an initial weight value of an extension neural network is determined according to training samples and the priori knowledge; (3) the extension neural network can be trained by the utilization of the training samples, if a training process is converged or the total error rate reaches a preset value, training is stopped, and a weight value vector, after the training, of the extension neural network is kept, and otherwise the training is continued; (4) the trained extension neural network is used for performing pattern recognition until recognition of all objects to be recognized is completed. According to the extension neural network pattern recognition method, under the common driving of the priori knowledge and the training samples, learning of the extension neural network is guided, training and learning of the extension neural network are completed, the learning burden of the extension neural network is relieved, the performance of the extension neural network is effectively improved, training time is shortened, and recognition accuracy is improved.

The invention discloses a method for designing a manifold regularization based semi-supervised classifier for dynamic vision. A training data source of the classifier under the environment of the dynamic vision is as follows: data with class information acquired by a user, positive class sample including objects, negative class sample not including objects, and data without class information randomly acquired by a computer in a video; continuity regular terms of the classifier on a sample data aggregation is defined by utilizing a local linear reconstruction coefficient of the data so as to make the utilization rate of the classifier on the data without the class information be improved. In addition, the definitions of a function complexity and function continuity regular terms are configured to be a form of 1 norm in an optimization problem of solving the coefficients of the classifier, thus the solution of the optimization problem of the classifier is a sparse solution of the coefficients of the classifier, namely the classifier obtained by training is also sparse. Therefore, the real time property of the classifier in a dynamic vision task is improved.

The invention relates to a digital signal modulation mode identification method, which comprises the following steps of: encoding an original data stream sequence of a transmitting end with a known modulation mode by utilizing a differential space-time encoding technology to obtain a code word sequence x (k), and estimating by utilizing autocorrelation matrixes J and x (k) to obtain a receiving signal sequence y (k); carrying out zero-forcing equalization technology processing on y (k) to obtain a received signal compensation matrix (shown in the specification), and calculating the received signal compensation matrix (shown in the specification) by utilizing different high-order cumulants (shown in the specification); and calculating the received signal compensation matrix (shown in the specification) by utilizing different high-order cumulants (shown in the specification). obtaining a received signal eigenvector matrix Cij according to the eigenvector of the received signal eigenvector matrix Cij and the eigenvector of the received signal eigenvector matrix Cij; And performing normalization processing on the Cij to obtain a normalized modulation mode feature vector CijF, and inputting the normalized modulation mode feature vector CijF and a deep learning network which is trained and built by the CijF and the category label pair. According to the system for identifying the modulation mode to be identified, a receiving signal sequence is collected at the receiving end of the system, a normalized modulation identification feature vector is obtained according to the method inthe above step and serves as the input of a trained classifier, the output of the classifier serves as the tag sequence of the system modulation mode, and modulation mode judgment is completed.

The invention provides an electrocardiosignal classifying method and system based on LRF-ELM and BLSTM. The electrocardiosignal classifying method includes the steps: acquiring electrocardiosignal data, preprocessing the electrocardiosignal data to obtain a dataset, and using the electrocardiosignal data in the dataset as input data of a neural network; using a LRF-ELM network as a feature extractor, learning spatial information in the electrocardiosignal data, and through three stacked random convolution and pooling processes, extracting feature data of different dimensions in the electrocardiosignal data; and after fusion, using the extracted feature data as input of a sequence learning stage, adopting a deep BLSTM network to carry out sequence learning, and finally outputting electrocardiosignal classifying results. According to the electrocardiosignal classifying method and system based on the LRF-ELM and the BLSTM, time information and spatial information of electrocardiosignals are taken into account at the same time, and therefore, not only can electrocardiosignal features be extracted efficiently and rapidly, but also good classification and identification properties are ensured.

The invention relates to a human body behavior recognition method of a non-local double-flow convolutional neural network model. Two shunt networks are improved on the basis of a double-flow convolutional neural network model; a non-local feature extraction module is added into the spatial flow CNN and the time flow CNN for extracting a more comprehensive and clearer feature map. According to themethod, the depth of the network is deepened to a certain extent, network over-fitting is effectively relieved, non-local features of a sample can be extracted, an input feature map is subjected to de-noising processing, and the problem of low recognition accuracy caused by reasons such as complex background environment, diverse human body behaviors and high action similarity in a behavior video is solved. According to the method, an A-softmax loss function is adopted for training in a loss layer; on the basis of a softmax function, m times of limitation is added to a classification angle, andthe weight W and bias b of a full connection layer are limited, so that the inter-class distance of samples is larger, the intra-class distance of the samples is smaller, better recognition precisionis obtained, and finally a deep learning model with higher identification capability is obtained.

The invention discloses a ship ISAR image recognition method based on optical image assistance, and relates to a ship ISAR image recognition method. The invention aims to solve the problem of low shipISAR image recognition accuracy in the prior art. The method comprises the following steps: 1, preprocessing ship ISAR images with known ship types in a database; 2, obtaining a trained derivative network Pix2pix; 3, generating a ship optical image from the ship ISAR image by using the trained derivation network Pix2pix, and merging the ship optical image and the ship ISAR image; obtaining a trained convolutional neural network; 4, preprocessing the ship ISAR image to be tested; 5, combining the obtained ship optical image with the ship ISAR image by using the trained derivation network Pix2pix; and 6, inputting the obtained composite image into the trained convolutional neural network to obtain a category. The method is applied to the field of radar target recognition.

The invention discloses a plaintext and ciphertext signal classification detection method for blind estimation of wireless signals. When the status of wireless plaintext and ciphertext is known, the features of plaintext signal and ciphertext signal are extracted respectively, and the features of plaintext signal and ciphertext signal are used as training set, and the detected phase statistical eigenvalues of wireless signal are used as test set, then the support vector machine is inputted to train and classify the features. Compared with the prior art, the positive effect of the invention isthat the invention proposes a novel wireless signal ciphertext security detection method aiming at the wireless network electromagnetic signal security detection problem, and solves the problem of blind identification classification based on the wireless signal modulation phase statistical characteristics under the conditions of non-demodulation and non-decoding. The invention has the advantages of good plaintext / ciphertext signal detection, classification and identification performance, high reliability, low cost and convenient use, and can efficiently meet the security analysis requirementsof ciphertext signals in various wireless network communication environments.

The invention discloses a method for detecting the geological condition of a tunnel face in the tunnel construction process. The method comprises the steps that based on a deep learning algorithm, a sample data set is used for training an image instance segmentation neural network to obtain an image instance segmentation model, and the sample data set comprises a plurality of rock slag sample images of different geological levels; each rock slag sample image is provided with a geological category label, and the contours of blocky rock slag and flaky rock slag are marked in the images. And the image instance segmentation model is called to analyze the to-be-identified rock slag image to obtain contour data segmented in the to-be-identified rock slag image corresponding to each rock slag in the solid slag and a probability value of the solid slag belonging to each geological level. And the content values of the blocky rock slag, the flaky rock slag and the rock powder in the solid muck are calculated according to the contour data, and the geological level of the tunneling tunnel face is determined in combination with the initial classification result, so that the geological analysis accuracy is not reduced while the defect of manual detection of the geological condition of TBM tunnel construction is overcome, and the intelligent degree of tunnel construction is improved.

The invention provides a digital communication signal modulation recognition method based on preprocessing noise reduction, which comprises the following steps: S1, constructing a modulation signal according to a carrier frequency, a code part rate and a sampling frequency, and carrying out noise adding processing; s2, carrying out noise reduction preprocessing on the signal after noise adding processing by using an adaptive filtering technology; s3, extracting wavelet transform features and high-order accumulation features of the signals after noise reduction preprocessing; s4, inputting the wavelet transform features and the high-order accumulation features into a BP neural network for network training; and S5, recognizing a modulation mode of an unknown signal by using the trained BP neural network. According to the method, a communication signal containing noise is preprocessed by using an adaptive filtering technology so as to improve the signal-to-noise ratio of the signal, then wavelet transform and high-order cumulant characteristics of the signal are extracted, and finally the wavelet transform and high-order cumulant characteristics are input into a neural network classifier, so that a better recognition effect can be obtained in a low signal-to-noise ratio environment.

The invention discloses an identification method and a system based on an integrated tree for false positive alarm signals in a ventilator. The identification method comprises the following steps: S1,data collection: collecting monitoring data of a patient from a ventilator and a monitor in a hospital; S2, data preprocessing: processing missing values and abnormal values in a dataset, conductingstandardization, and generating rules for identifying false positive alarm signals; S3, feature extraction: sorting features by using a random forest, and selecting the features with good recognitioncapability; and S4, false positive alarm signal identification: establishing a false positive alarm signal identification method for the ventilator and the monitor. Experimental results show that themethod of the present invention has excellent identification performance for false positive alarm signals and robust identification effect.

The invention relates to a face recognition method based on combined sparse representation and single hidden layer neural network technology. Compared with the prior art, the method solves the problem of a time-consuming recognition method or unstable classification performance. The method provided by the invention comprises the steps that face image acquisition and detection are carried out, and the range of five sense organs of a face is determined in a test frame image extracted from a video; image preprocessing is carried out, and illumination or noise interference eliminating is carried out on the test frame image; feature extraction is carried out, face feature extraction is carried out on the test frame image; and classification recognition is carried out, and extracted face feature information is subjected to search classification matching with a feature template stored in a database to acquire a final classification recognition result. According to the invention, a sparse representation method is combined with the single hidden layer feedforward neural network; the recognition speed is fast; and a good classification recognition performance is kept.

The invention discloses an image classification and identification method based on the regional bicubic interpolation technology, and belongs to the field of image processing. An existing image processing method has the problem that the image classification identification rate is reduced due to image distortion caused by scale transformation. According to the image classification and identification method based on the regional bicubic interpolation technology, the method comprises the steps of extracting the features and K-of a training sample through SIFT; performing means clustering statistics on the word frequency, and training a classifier model after training data is obtained in the classification step; subjecting the images of a test sample to identical zooming-in and zooming-out changes through regional bicubic interpolation, and then carrying out retrieval and classification through the BOF algorithm. According to the invention, the classification and identification accuracy ofdistorted images caused by zooming can be improved.

The invention provides a human body action recognition method, and relates to the field of action recognition. The method comprises the steps: obtaining feature data of articulation points, and establishing articulation point set distance features, geometrical features and motion features; modeling the spatial and temporal features of the action sequence from multiple angles, and modeling the timesequence information of the action sequence by adopting a one-dimensional time convolution network; through time-space characteristics and time sequence information, carrying out the human body action classification and identification; through designing a plurality of characteristic representations, describing the space and geometric information of different joint points in the same frame, describing the time motion information of joint points between adjacent frames, and better modeling the space-time characteristics of human body motion; meanwhile, employing a one-dimensional time convolution network for modeling time sequence information, carrying out the human body motion classification in combination with constructed characteristics, and obtaining an ideal classification and recognition effect.

The invention discloses an SAR image target recognition method based on sparse representation and a cascade dictionary. The method comprises the steps: firstly carrying out the template training of asample image based on a maximum extended average correlation height filter; secondly, extracting monogenic features of the template image, namely monogenic amplitude representing signal energy, monogenic phase representing signal structure information and monogenic orientation representing signal geometric information, constructing sub-dictionaries by the three features with complementary properties, and cascading the plurality of sub-dictionaries by using each sub-dictionary as a classifier; and finally, SAR image target classification is realized based on a classification mechanism with themaximum sparse representation coefficient energy and the minimum reconstruction error, and a good classification and recognition effect can be realized.

The invention discloses an ISAR ship target image domain enhanced recognition method based on a loop generative adversarial network, and relates to the field of radar target recognition. The objectiveof the invention is to solve the problem of low accuracy of ship type identification in the prior art. The method comprises the steps: preprocessing ship original ISAR images and original optical images of various ship types stored in a database to obtain an original ISAR image data set and an original optical image data set; taking the original ISAR image data set and the original optical imagedata set as a training set, and training the training set by adopting a loop generative adversarial network to obtain a trained loop generative adversarial network; preprocessing a to-be-tested ship ISAR image or optical image to obtain a to-be-tested ISAR image or optical image; analyzing the to-be-tested ISAR image or optical image by adopting the trained loop generative adversarial network to obtain a new image with optical characteristics and ISAM characteristics at the same time; and analyzing the new image by adopting a convolutional neural network to obtain a ship type. The method is used to identify images.

The invention discloses a two-phase coding radar signal distance super-resolution method based on target sparsity. The method comprises the following steps: 1) carrying out matched filtering processing on a two-phase coding echo signal of a radar target; 2) establishing a distance dimension sparsity model of the target echo signal; and 3) resolving the maximum posterior probability estimation of the sparse amplitude vector to obtain a target signal amplitude, and obtaining a target distance and realizing distance super-resolution. Under the condition that the bandwidths of the two-phase coded signals are the same, the Rayleigh limit of conventional pulse pressure is broken through, and distance super-resolution is achieved; and radar target detection, imaging and classification identification performance is improved by distance super-resolution.

The invention provides a liver tumor MVI prediction method based on a triple network. The method comprises the following steps: dividing liver MRI image data into a training set and a test set; selecting three samples from the training set to form a triple sample; constructing a triple network model, wherein the triple network model is composed of three paths of convolutional neural networks with the same structure, and the triple network model receives one triple sample as input each time; converting an input triple sample into a vector in an embedded layer space through the triple network model, calculating triple loss, calculating cross entropy loss by using a convolutional neural network in the triple network model, and fusing the triple loss and the cross entropy loss to serve as a target loss function for training the triple network model; and training the triple network model by using the training set to obtain an optimal triple network model, and performing liver tumor MVI prediction on the liver MRI image data in the test set by using the trained triple network model.

The invention relates to an industrial-grade intelligent surface defect detection method. By constructing and training a twin generation adversarial network GAN, the input image is repaired into a "normal sample" through an improved GAN network, and the output is compared with a manually labeled positive sample through the twin The CNN network is compared, and the difference is a defect. The twin generation adversarial network designed by the present invention does not require a large number of samples or data amplification, which can solve the problem of small sample sizes of common industrial products and reduce the occurrence of over-fitting phenomena caused by deep learning with few samples and zero samples. , making it possible to detect defects in products with a small number of defective samples and in new product development. The cross alignment loss function CA and the distribution alignment loss function DA are used to strengthen the relationship between the outputs of the two networks and achieve better classification and recognition results. The model training speed is improved through the Attention mechanism and hardware GPU, making rapid industrial deployment possible.

The invention discloses a plaintext signal classification and detection method for blind estimation of wireless signals. In the case of known wireless plaintext and ciphertext states, the plaintext signal features and ciphertext signal features are respectively extracted, and the plaintext signal features are extracted. The ciphertext signal features are used as the training set, and the detected wireless signal phase statistical feature values ​​are used as the test set, which is input into the support vector machine for feature training and classification judgment. Compared with the prior art, the positive effect of the present invention is that the present invention proposes a novel wireless signal ciphertext security detection method aiming at the problem of wireless network electromagnetic signal security detection, which solves the problem of non-demodulation and non-decoding conditions. The blind recognition classification problem based on the statistical characteristics of wireless signal modulation phase. The invention has good detection and classification recognition performance for plaintext\ciphertext signals, high reliability, low cost, convenient use, and can efficiently meet the security analysis requirements of ciphertext signals in various wireless network communication environments.

The invention discloses a PCB defect image recognition method based on an improved convolutional neural network. In order to overcome the defects of an existing ResNet50 deep convolutional neural network model, a novel CNN module named Res2Net is introduced, and a residual connection structure and an activation function are changed, so that the multi-layer nonlinear expansion capability of the network is improved. Based on an improved ResNet50 model, a DenseNet169 convolutional neural network is fused, fusion is carried out based on image features extracted by multiple models, a fusion feature output network structure is improved, and a multi-model fused convolutional neural network framework for PCB defect image recognition is established. The method can identify different types of PCB defect images, has the characteristics of high identification accuracy and sensitivity compared with a single model, and can realize automatic and intelligent identification of PCB defect types.

The invention provides a polarization characteristic multi-scale pooling classification algorithm for a complex vehicle road environment, and classification of image targets in the complex vehicle road environment is realized. Firstly, vehicle road environment medium conditions are analyzed, and a high-quality imaging mode in a complex vehicle road environment is explored; secondly, a polarizationimaging scheme is designed based on a simulation experiment result, and a three-channel imaging system is assembled and calibrated. finally, a multi-scale pooling deep semantic recognition algorithmis proposed to realize recognition and classification of vehicle and road environment targets. Experimental results show that the semantic classification recognition effect of a complex scene can be effectively improved, and reliable technical guarantee is provided for safe auxiliary driving visual perception of vehicles in a complex vehicle road environment.

The invention discloses a millimeter wave fuse chaff interference identification method based on distance image feature extraction. The method is good in stability, high in accuracy, high in processing efficiency and good in classification effect. The method comprises the following steps: (10) target distance image alignment: processing a fuse target echo signal, and obtaining an aligned target distance image by utilizing global minimum entropy correction; (20) obtaining a chaff cloud distance image: processing the chaff interference signal based on a chaff cloud dynamic diffusion model to obtain a chaff cloud one-dimensional distance image; (30) feature calculation: analyzing and aligning the target distance image and the chaff cloud one-dimensional distance image by using an entropy algorithm to obtain a waveform entropy value of the target / chaff distance image, calculating distance image correlation and a scattering intensity ratio of the target distance image and the chaff cloud one-dimensional distance image, and taking the distance image correlation and the scattering intensity ratio as a feature parameter set; (40) feature extraction: performing clustering analysis on the feature parameter set by using an FCM algorithm to obtain clustering distribution of a target and interference; and (50) judgment and identification: setting a threshold value, and judging the type of the fuse echo signal.

The invention belongs to the technical field of cerebral stroke patient rehabilitation training, and particularly relates to a cerebral stroke patient-oriented cross-subject rehabilitation training method, which comprises the following operation steps of: 1, collecting motor imagery electroencephalogram data of a healthy person during rehabilitation training simulation by using electroencephalogram acquisition equipment so as to obtain the electroencephalogram data of the healthy person, 2, the obtained healthy human electroencephalogram data and Gaussian noise are generated into electroencephalogram data through a generator, the electroencephalogram data can be expanded, the problems that the electroencephalogram data is small in data amount, large in collection difficulty and high in cost are solved, the calibration time when a new user uses a rehabilitation system is shortened, and the usability of the system is improved; and the electroencephalogram mode of the motor imagery of the healthy person is utilized to help to identify the motor imagery task of the stroke patient, and the identification precision is improved, so that the generalization of the rehabilitation system is improved, and the rehabilitation system is easier to use for a new user.

The invention provides an electrocardiosignal classifying method and system based on LRF-ELM and BLSTM. The electrocardiosignal classifying method includes the steps: acquiring electrocardiosignal data, preprocessing the electrocardiosignal data to obtain a dataset, and using the electrocardiosignal data in the dataset as input data of a neural network; using a LRF-ELM network as a feature extractor, learning spatial information in the electrocardiosignal data, and through three stacked random convolution and pooling processes, extracting feature data of different dimensions in the electrocardiosignal data; and after fusion, using the extracted feature data as input of a sequence learning stage, adopting a deep BLSTM network to carry out sequence learning, and finally outputting electrocardiosignal classifying results. According to the electrocardiosignal classifying method and system based on the LRF-ELM and the BLSTM, time information and spatial information of electrocardiosignals are taken into account at the same time, and therefore, not only can electrocardiosignal features be extracted efficiently and rapidly, but also good classification and identification properties are ensured.