35results about How to "To solve the classification accuracy is not high" patented technology

The present invention discloses a feature extraction and state recognition method for one-dimensional physiological signal based on depth learning. The method comprises: establishing a feature extraction and state recognition analysis model DBN of a on-dimensional physiological signal based on depth learning, wherein the DBN model adopts a "pre-training+fine-tuning" training process, and in a pre-training stage, a first RBM is trained firstly and then a well-trained node is used as an input of a second RBM, and then the second RBM is trained, and so forth; and after training of all RBMs is finished, using a BP algorithm to fin-tune a network, and finally inputting an eigenvector output by the DBN into a Softmax classifier, and determining a state of an individual that is incorporated into the one-dimensional physiological signal. The method provided by the present invention effectively solves the problem that in the conventional one-dimensional physiological signal classification process, feature inputs need to be selected manually so that classification precision is low; and through non-linear mapping of the deep confidence network, highly-separable features/feature combinations are automatically obtained for classification, and a better classification effect can be obtained by keeping optimizing the structure of the network.

The invention discloses a Chinese text sorting method based on correlation study between sorts. The method comprises the following steps of: firstly, dividing words of a document and performing rough selection on characteristics by computing word frequencies; secondly, further determining representative word items according to discrimination indexes between the word items and sorts so as to form characteristic word items which are finely selected; thirdly, training the document to be expressed by a tfidf weight and a discrimination index weight according to an index of the characteristic word items; fourthly, establishing a group of two-sort sorters corresponding to different projection vectors and training to obtain a code array expressing the correlation between two-sort sorters; and finally, projecting a multi-vector expression of a new document to all the two-sort sorters, introducing the code array, computing the similarity between each sort and the document, and outputting the maximum of the similarity as a sort judging result of the new document. The new document is sorted based on a correlation studying result between the sorts, and the running efficiency of an algorithm is improved on the premise of ensuring the sorting performance.

The invention discloses a hyperspectral image classification method based on compression spectrum clustering integration. The hyperspectral image classification method comprises a classification process including the steps that (1) image characteristic sets of hyperspectral images are obtained; (2) an image characteristic set sub space subjected to dimension reduction is obtained; (3) a plurality of hyperspectral image dividing results are obtained; (4) the final hyperspectral image dividing result is obtained; (5) the hyperspectral image classification result is obtained; (6) the accurate classification of the hyperspectral image is obtained. Compared with the prior art, the hyperspectral image classification method has the advantages that the defect that a k-means algorithm adopted in the classical spectral clustering is sensitive to initialization is overcome; the characteristic dimension of the hyperspectral images is reduced; meanwhile, the classification precision is obviously improved, and the dividing effect is good.

The invention discloses a small sample PolSAR image classification method based on fuzzy label semantic prior. The method comprises the steps of preparing a PolSAR image to be classified; obtaining real polarization characteristics as input data of the network; obtaining a sampling matrix for recording the position of the training sample and a sampling label matrix for recording pixel label information at the corresponding position; utilizing the sampling label matrix to initialize and classify to build a full convolutional network FCN; sending the real-number input data, the sampling matrix,the sampling label matrix and the classification matrix to a built full convolutional network FCN for training; updating the classification matrix by utilizing the prediction result of the FCN, the sampling matrix, the sampling label matrix and the current state of the classification matrix; repeating the operation until the maximum number of iterations is met; outputting the final classificationmatrix; and calculating classification accuracy and a classification result graph to complete image classification. According to the invention, alternate iteration training is carried out on the deepfull convolution network parameters and the label category variables, and the problem of low PolSAR classification precision under a small sample problem is solved.

The invention discloses a magnetotelluric signal denoising method based on noise discrimination, which comprises the following steps: calculating approximate entropy and LZ complexity of each electromagnetic signal sample; using the approximate entropy, LZ complexity and class value of each electromagnetic signal sample to train the preset classification model to get the noise discrimination classification model; acquiring magnetotelluric signals to be processed, and performing noise screening on the magnetotelluric signals to be processed according to the noise screening classification modelto obtain electromagnetic signal segments with non-strong interference and electromagnetic signal segments with strong interference; combining the empirical mode decomposition of complementary set andwavelet threshold method, the strong disturbance electromagnetic signal being suppressed; the reconstructed magnetotelluric signal being obtained by combining the de-noising suppressed electromagnetic signal with the non-strong disturbance electromagnetic signal. The method of the invention can more accurately discriminate the data segments with strong interference and non-strong noise interference, retain the real magnetotelluric signal, and improve the denoising effect of the magnetotelluric signal.

The invention discloses a polarization SAR terrain classification method based on deep learning and distance metric learning. The polarization SAR terrain classification method comprises the realization steps that (1) images are inputted; (2) filtering is performed; (3) features are extracted; (4) training samples and test samples are selected; (5) a stacked sparse auto-encoder is trained so that the deep features of a training sample set and the deep features of a test sample set are obtained; (6) a distance metric learning classifier is trained so that a classification result is obtained; (7) the classification result is colored; and (8) the colored classification result graph is outputted. The images are classified by using the polarization SAR terrain classification method based on deep learning and distance metric learning so that feature extraction is relatively comprehensive and reasonable, the classification result is more consistent with real terrains, time complexity is reduced and classification precision is enhanced.

The invention provides a direction changing difference expansion and synchronous embedding reversible watermark embedding and extraction method. Embedding capacity and embedding data are enabled to be unrelated by adopting direction changing difference expansion and the threshold is enabled to select the maximum embedding capacity; and the classification accuracy and the maximum embedding capacity are enhanced by adopting gradient and direct classification for avoiding normalized gradient and reduction of the classification accuracy. A compressed location graph is applied so that large consumption of the embedding capacity and excessive additional data caused by recording overflow pixel location of a generic difference expansion and local complexity reversible watermark method can be avoided. The embedding data are ensured to be completely reversible by giving the embedding strategy of synchronous addition of backup data for avoiding irreversibility caused by direct embedding of the additional data after embedding of load data, and an embedding parameter selection method based on ordering and enumeration is also given so as to reduce computational complexity. Compared with the reversible watermark method based on generic difference expansion and local complexity, the method is completely reversible and parameter selection time is greatly reduced so that the method has larger maximum embedding capacity.

The invention relates to a set opening trace image classification method and system. Trace images are automatically classified. The method judges whether to-be-classified images belong to some type in a trace image library or not or a newly increased type in a multi-layer cascaded manner. The method and system provided by the invention have the characteristics of being fast in speed, less in artificial participation and relatively accurate and comprehensive in result, and can be used for automatically expanding the image library and serially and parallelly connecting cases by classifying traces on the scene, thereby providing a great help for police investigation. Therefore, the method and system provided by the invention can be widely used for the field of trace classification.

The invention provides a realization method for an analysis model supporting massive long text data classification, and belongs to the technical field of big data text analysis. According to the method, standard word segmentation in an HanLP word segmentation tool and an improved CHI algorithm are adopted, so that on one hand, the dimension of a word vector space of each article during text classification is effectively reduced, the time complexity of text classification computing is lowered, the algorithm efficiency is improved, and the performance demand during massive long text classification under the background of big data is met; and meanwhile, the situation of reduced classification accuracy due to reduced vector space dimension is reduced to the maximum extent. A barrier between a text and a vector can be effectively eliminated by adopting a TFIDF algorithm, and finally the text can be accurately subjected to better training by adopting a Naive Bayesian classification algorithm, so that accurate classification of the long text is realized. The method can effectively solve the problem of contradictoriness of a performance index and an accuracy index of long text classification in a big data environment, and has a wide application prospect.

The invention provides a full-duplex end-to-end automatic encoder communication system and an anti-eavesdropping method thereof, the full-duplex end-to-end automatic encoder communication system comprises a transmitter, a legal receiver and an eavesdropper, confidential information transmitted by the transmitter is transmitted to a wireless channel after being encoded by a multilayer neural network and a normalization layer; additive white Gaussian noise and an anti-disturbance signal in a wireless channel are added by a noise layer, and a confidential signal is decoded from a signal reachinga legal receiver through a multi-layer neural network and a softmax activation layer. Classification precision of the tapping network can be greatly reduced with relatively low power, so that the biterror rate of the tapping is greatly increased, and safe transmission of information between legal nodes is effectively protected; according to the full-duplex legal receiver, on one hand, disturbanceof a loop-back channel is reduced through an interference elimination method, and on the other hand, training is carried out through an adversarial training method, so that the legal receiver can resist self-disturbance from the loop-back channel, and therefore it is guaranteed that the bit error rate of the legal receiver cannot be increased due to self-disturbance.

The invention discloses a language model fine-tuning method for low-resource adhesive language text classification, and relates to the technical field of language processing; a low-noise fine-tuning data set is constructed through morphological analysis and stem extraction; a cross-language pre-training model is fine-tuned on the data set to obtain a fine-tuning result; a meaningful and easy-to-use feature extractor is provided for a downstream text classification task, related semantic and syntactic information is better selected from a pre-trained language model, and these features are used for the downstream text classification task.

The invention discloses a hyperspectral image classification method and device based on correlation coefficients and joint sparse representation, and belongs to the technical field of image classification. The hyperspectral image classification method comprises the steps: calculating the correlation between two variables through employing the correlation coefficients in combination with the similarity of spectrums and the local space consistency, so as to calculate the correlation between a pixel and each class; calculating a sparsest matrix by using an SOMP algorithm, and calculating a residual error between a pixel and each class by using the matrix; and moreover, constraining the classification function model by using the joint sparse representation and the correlation coefficient, andintroducing a parameter to balance the weights of the joint sparse representation and the correlation coefficient, so that the adaptability and robustness of the model are promoted, and compared withother methods, the hyperspectral image classification method is higher in accuracy and more stable in performance.

The invention discloses a hyperspectral image spectral space classification method based on class characteristic iterative random sampling. The method comprises the following steps: calculating classcharacteristic criteria of each class according to hyperspectral image data and ground object class label information; iteratively calculating the number of training samples which should be allocatedto each category during classification according to a category feature criterion; calculating an average spectral characteristic and an autocorrelation matrix of a target ground object according to the number of the distributed training samples of each type and the parameter information of the target ground object sample set; calculating abundance information corresponding to each pixel in the hyperspectral image for the ground object through a constraint energy minimization method according to the average spectral characteristic of each ground object and the inverse matrix of the autocorrelation matrix; and calculating a similarity coefficient of two adjacent iterative classification results according to the classification image information, judging whether the similarity coefficient meets an iterative classification cut-off condition, and carrying out iterative classification on the fused and updated data.

The invention belongs to the technical field of image processing and computer vision, and discloses an image quality evaluation method and system. The method comprises the following steps: image preprocessing: carrying out equal-proportion scaling and normalization on an image, and carrying out default value filling on a lacking part; image effective position prediction and image filtering: determining a certificate image position and a certificate category based on a pixel-level segmentation-classification technology model, and filtering the image by setting a threshold value; correcting the to-be-evaluated image: segmenting the certificate image according to the certificate image position information and correcting the certificate image; multi-dimensional quality evaluation: performing quality evaluation and scoring from five dimensions of type, integrity, definition, light spot and PS judgment based on deep learning and a convolutional neural network technology model; and outputting a result in a structured manner. According to the method, the to-be-evaluated certificate image is effectively distinguished by adopting a pixel-level image segmentation-classification technology, multi-dimensional evaluation scoring is performed in combination with deep learning, a convolutional neural network technology and a traditional method, an extracted prediction result is efficient and accurate, and an evaluation result is accurate.

The invention discloses a texture description method and system based on an improved local binary pattern. According to the invention, on one hand, the method includes extracting all equivalent modesin an original image; extracting all non-equivalent modes in the original image at the same time, taking a part of non-equivalent modes with high occurrence frequency from the non-equivalent modes asdominant non-equivalent modes based on the occurrence frequency of each type of non-equivalent modes, and constructing a mixed mode by utilizing all the equivalent modes and all the non-equivalent modes; on the other hand, the method includes selecting a rotary encoder or a natural encoder for hybrid encoding according to the image noise condition, and when the image is in a normal state, selecting the rotary encoder; and when the image is in a light alarm state, selecting a natural encoder. According to the method, texture description information in a non-equivalent mode is mined, the problemthat the texture description capability is reduced after a traditional LBP ignores the non-equivalent mode is solved, meanwhile, two coding modes which are freely switched according to the image noise degree are provided, and the anti-noise performance of the texture description process is improved.

The invention discloses a hyperspectral classification method based on a deep feedforward network. The method includes: utilizing a training sample allocation algorithm to calculate the number of training samples needing to be allocated to each category during classification; in each layer of classification network, generating a training sample data set for training a classifier by adopting a fixed training sample selection mode or a random training sample selection mode according to the number of the distributed training samples of each category; selecting a support vector machine or a convolutional neural network to perform initial classification on the images to obtain an initial classification result; extracting spatial feature information of the classification graph by using an edge preserving filter, and re-classifying the spatial feature information by using a trained support vector machine; and judging whether a stop condition is met or not, and if not, entering a lower-layer network for classification in a feedforward mode until an optimal classification result is finally obtained. According to the classification framework, through a series of spatial filters and feedforward operation, spatial feature information of the hyperspectral image is effectively mined, and an initial classification result is improved.