47results about How to "Sparse" patented technology

The invention discloses a sparse representation based short-voice speaker recognition method, which belongs to the technical field of voice signal processing and pattern recognition, and aims to solve the problem that the existing method is low in recognition rate under limited voice data conditions. The method mainly comprises the following steps: (1) pretreating all voice samples, and then extracting Mel-frequency cepstral coefficients and first-order difference coefficients thereof as characteristic; (2) training a gaussian background model by a background voice library, and extracting gaussian supervectors as secondary characteristics; (3) arranging the gaussian supervectors for training voice samples together so as to form a dictionary; and (4) solving an expression coefficient by using a sparse solving algorithm, reconstructing signals, and determining a recognition result according to a minimized residual error. According to the invention, the gaussian supervectors obtained through self-adaption can greatly relieve the problem that the personality characteristics of a speaker are expressed insufficiently due to limited voice data; through carrying out classification by using sparsely represented reconstructed residual errors, a speaker model mismatch problem caused by mismatched semantic information can be handled.

The invention relates to wave arrival direction estimation method under an array error. First the method utilizes sparsity of a matrix, expresses wave arrival direction estimation under the array error by solving an optimization problem, and then converts the optimization problem into an iterative algorithm to realize. Compared with an existing wave arrival direction estimation algorithm under an array error, the estimation method proposed by the invention has remarkably improved estimation performance under the condition of equivalent complexity, and has good application prospects in the field of direction estimation.

The invention discloses a local binary CNN processing method, a device, a storage medium and a processor, wherein the method comprises the steps of training a first convolutional neural network according to a preset data set, and obtaining a second convolutional neural network; replacing all convolutional layer units in the first convolutional neural network by a local binary convolutional unit, thereby obtaining a third convolutional neural network; replacing objective total connecting layers in the third convolutional neural network by a grouping connecting layer, thereby obtaining a fourth convolutional neural network, wherein the objective total connecting layers are all total connecting layers except for a bottom classification layer in the third convolutional neural network; initializing the fourth convolutional neural network, thereby obtaining a fifth convolutional neural network; and training the fifth convolutional neural network based on the second convolutional network and a preset data set, thereby obtaining an objective convolutional neural network. The local binary CNN processing method, the device, the storage medium and the processor settle a technical problem of relatively low operation efficiency in the local binary convolutional neural network in prior art.

The invention discloses a denoising method of a low-rank image, which is applied to the reconstruction of an image containing noise. By analyzing the correlation characteristics of the low-rank matrix, the denoising problem of the image is modeled as a non-linear constraint problem, and then the specific iterative steps are deduced by using an alternating direction multiplier method, thereby obtaining the image after denoising. The invention uses the truncated kernel norm to more accurately approximate the rank of the matrix and avoids the result deviation caused by the large singular value. In addition, the Frobenius norm is added to the model as a regular term, the Frobenius norm and the truncated kernel norm form an elastic network about the singular value, so that the final result is sparse and stable, so that better denoising effect can be achieved.

The invention discloses an embedded low-power convolution neural network method; for the two adjacent layers of the conv layer and the conv+concat layer, the idea of ​​inception in GoogLeNet is adopted, and the volume of 1×1 and 3×3 The product kernels are arranged in parallel, and features on different scales are collected separately, and then the feature maps calculated by the two convolution kernels are spliced. For the conv+concat layer, the idea of ​​​​ResNet is adopted. The splicing results of this network in the conv+concat layer The element-by-element addition operation of short cut connection was performed, and each layer was pre-pruned separately. The cut neurons accounted for 50% of the total, and the impact of each layer on the overall accuracy was tested. Reduce the convolution size of each layer of the multi-layer network, reduce the number of convolution kernels and the network has a certain degree of sparsity, so the number of multiplications required for each layer is very small. In specific calculations, it can be converted to have CSR The product of the sparse matrix and the dense matrix in the storage format, in order to achieve the purpose of small memory usage, fast calculation speed and high precision.

The invention discloses a domain transfer extreme learning machine method based manifold regularization and norm regularization. On the basis of a traditional extreme learning machine, the thought of semi-supervised learning and transfer learning is introduced, and a novel extreme learning machine model is built and consists of three parts: a manifold regularization term capable of excavating geometric distribution shapes of data samples with tags and without tags to realize semi-supervised learning; a loss function term considering error minimization of source domain data and target domain data to realize transfer learning; and norm regularizers constraining weight space. The domain transfer extreme learning machine method provide by the invention is combined with the source domain to process the problem of prediction of the target domain, thereby increasing the generalization capability and range of application of the extreme learning machine. Introduction of the manifold regularization term also enables the method proposed by the invention to still maintain a relatively good learning effect when data with tags are little, the restriction that a traditional machine learning method requires a large amount of data with tags is overcome, and the accuracy and robustness of prediction are also improved.

The present invention provides an ACO-OFDM wireless optical communication system channel estimation method based on compressed sensing, and belongs to the field of intensity modulation optical communication. The channel estimation method is used to solve the problem that during pilot based channel estimation, a pilot symbol causes a bandwidth utilization rate to be reduced. According to the method provided by the present invention, channel estimation technology based on the compressed sensing is combined with the ACO-OFDM system. Firstly, the pilot is enabled to have Hermitian symmetry, a pilot allocation scheme is optimized by minimizing a DFT sub-matrix cross-correlation squared sum, and an ACO-OFDM system pilot allocation optimization algorithm is provided. Then, a channel response is estimated by using an improved variable step sparse degree adaptive matching tracking algorithm, the improved algorithm can improve estimation accuracy and calculation speed. Finally, an iterative correction algorithm is used to further improve the estimation accuracy. Compared with the traditional pilot based channel estimation method, the method can effectively reduce the number of pilot symbols, and improve the system bandwidth utilization rate.

The invention relates to a low-rank decomposition and space spectrum constraint-based hyperspectral image time domain change feature extraction method. The method mainly comprises the following stepsof: calculating a difference image between two hyperspectral remote sensing images acquired at a same place and at different times; establishing a low-rank matrix decomposition model with a space spectrum constraint by utilizing internal data structural characteristics of the difference image; and solving each component of the model through an alternate iteration manner so as to extract a time domain change feature. According to the method, the low-rank matrix decomposition model with the space spectrum constraint and a solution algorithm of the model are disclosed by sufficiently utilizing internal structures of data, so that the time domain change feature is effectively extracted, multiple forms of noises are removed, real changes are strengthened and false changes caused by noises are suppressed, thereby improving the change detection precision of time domain change hyperspectral remote sensing images.

The invention provides a face expression recognition method based on a convolution neural network, which relates to the technical field of face expression recognition. At first, the method collects facial expression pictures and downloads the FER-2013 face database and CK + face database; the images are divided into a training set and a test set, then the collected images are preprocessed to obtain the 96*96 size gray-scale images, and the convolution neural network model is established; the training set is used for training and the error between the actual output result of training and the label value is calculated, and the difference value is transmitted from top to bottom through the back propagation algorithm, and the weight value is updated by the weight value update formula. After the training, the network model is saved, the image of the test set is inputted into the training model, and the recognition rate is calculated. In the invention, the face expression recognition methodis improved, the convergence speed of the model is improved, the recognition efficiency is improved, the accuracy rate of the convolution neural network is changed, and the face expression recognitionefficiency is improved to a certain extent.

The invention discloses a nonorthogonal multi-carrier transmission method. The transmission method is based on a compressive sensing theory, and comprises the following steps of: expanding mutually orthogonal sub-carriers in an OFDM system into nonorthogonal sub-carriers with different sampling start time on a plurality of time domains; combining transmission of the plurality of nonorthogonal sub-carriers with compressive sensing, thus allowing signals mapped to the nonorthogonal sub-carriers to be sparse, and acquiring the mapped signals; and recovering the original signals at a receiving end via a restructing algorithm. According to the method provided by the invention, a total spectral bandwidth occupied by the carriers is reduced by nonorthogonality among the carriers at a same symbol rate, and spectrum efficiency of the nonorthogonal multi-carrier system is improved.

The invention discloses a compression acceleration method of a deep convolutional neural network for target detection. The method comprises the following steps: constructing and training a deep convolutional neural network for target detection; carrying out quantitative test on all weight values in the deep convolutional neural network and activation values of all layers except the last layer after passing through an activation function, testing the detection performance loss condition of the network from small to large in quantitative step size, and selecting the maximum quantitative step size in a set loss range; determining a truncation range of a weight value and an activation value in the neural network by utilizing the quantization step length, limiting the neural network and training the network; and truncating and quantifying the deep convolutional neural network, and compiling a forward code. According to the method, the quantization technology is adopted to reduce the networkstorage capacity, 32-bit floating-point number operation in the network is converted into 8-bit integer operation, and meanwhile, the sparsity of the network is utilized to convert the layer meetingthe sparsity condition in the network into sparse matrix operation, so that the purpose of compressing and accelerating the deep convolutional neural network is achieved.

The invention discloses a dense multipath signal angle estimation method based on impulse response compression sensing. A pseudo random sequence with the length of N as a baseband detection signal, aspatial channel impulse response model is determined, receiving array antennas that have the M same array elements and are arranged uniformly and linearly are generated, and a guide vector of the antenna array under a first path is determined; for receiving signals of from a first antenna element to an M antenna element, sliding correlation is carried out on the receiving signasl and a standard local pseudo-random sequence to obtain an observation impulse response, discretization processing is carried out to obtain an observation impulse response matrix, impulse response covariance matrixes ofeffective paths from the first one to the L one of the spatial channel are calculated respectively, a redundant dictionary for sparse reconstruction is constructed, and sparse vectors are calculatedto form an angular spatial spectrum, wherein the angle corresponding to the large value is the coherent path direct-of-arrival angle in the l-th path. Therefore, the direct-of-arrival angle estimationof lots of multi-path signals is realized; the estimation accuracy and the angular resolution are high; and the coherent paths are distinguished.

The invention discloses an image category detection method, and the method comprises the steps: obtaining a training image set, and constructing a training sample according to the image features of each training image in the training image set; training a classification model by using the training sample, wherein the classification model is a model constructed by a sparse Fisher support vector machine classification algorithm, and the sparse Fisher support vector machine is a Fisher support vector machine based on an L1 norm; and inputting a to-be-detected image into the trained classificationmodel to obtain the image type of the to-be-detected image. The image classification accuracy can be improved. The invention further discloses an image category detection system, electronic equipmentand a storage medium, which have the above beneficial effects.

The invention discloses a multi-path number estimation method based on a channel observation impulse response model. A pseudo-random sequence of length N is selected to be used as a baseband detectionsignal, the spatial channel impulse response model is determined, a receiving array antenna which has M identical array elements and is in uniform linear arrangement is generated, and a steering vector of an antenna array is determined; sliding correlation is carried out on received signals of first-to-M-th receiving antenna oscillators and the standard local pseudo-random sequence to obtain observation impulse response, discretization is carried out to obtain an observation impulse response matrix, multi-path number estimation is carried out on the observation impulse response matrix, inspection and estimation are carried out on first-to-K-th valid paths and coherence paths of spatial channels, and the multi-path number of the entire received signals and a corresponding correlation structure are obtained. According to the method, number estimation of the massive multi-path signals is realized, estimation accuracy is high, and the coherence paths can be distinguished.

The invention relates to a method for face recognition by adopting two-dimensional non-negative sparse partial least squares. The method comprises the following steps: firstly constructing a category matrix of a face training sample set, then constructing an objective function for realizing minimum loss of the amount of information in the projection of the face training sample set and the category matrix, afterwards adding non-negativity constraint and sparsity constraint into the objective function to obtain a convergent non-negative basis matrix, projecting the face training sample set on the basis matrix to obtain a test sample coefficient matrix, also carrying out the operation on a test sample to obtain a coefficient matrix of the test sample, determining that certain element matrices in the coefficient matrix of the test sample and the test sample coefficient matrix fall into the same category by using a nearest neighbor strategy, and then considering that a person on the test sample corresponding to the coefficient matrix of the test sample is the same person on the training sample corresponding to the element matrix. The method is high in recognition rate and robustness, and as only the basis matrix needs iterative solution, the operation is simplified, the time complexity is reduced, and the recognition speed is high.

The invention discloses an event camera target recognition method based on a self-attention mechanism. The event camera target recognition method comprises the following steps: S1, initializing an event camera; s2, completing a data acquisition task by using the initialized event camera; s3, performing image conversion on the collected event camera data, so that the event camera data can be used for a target identification task; s4, performing feature extraction on the event camera data after image conversion by using the trained network to obtain depth features of the target; s5, inputting the extracted depth features into a self-attention mechanism model for self-attention calculation to obtain target types included in the event camera data at the current moment; and S6, performing result output on the features subjected to self-attention calculation, and taking a result with the highest confidence coefficient as a final result to be output. The method solves the technical problems that current event camera training data is insufficient, and a traditional deep learning method is weak in global perception capability of event camera data.

The embodiment of the invention discloses a method and a device for detecting abnormal articles in a scene, and the method comprises the steps: respectively inputting a scene detection graph and a scene standard graph into corresponding sub-networks of a twin network, and respectively obtaining a feature graph of the scene detection graph and a feature graph of the scene standard graph; fusing thefeature map of the scene detection map and the feature map of the scene standard map to obtain a common feature map; calculating by adopting a plurality of hole convolution and the common feature maprespectively to obtain sub-feature maps of multiple types of articles in a scene encircled by a target frame; screening out sub-feature maps with abnormal articles from the sub-feature maps of the multiple types of articles in the scene through a full-connection network, calculating coordinates of the sub-feature maps with the abnormal articles, and circling the abnormal articles by using the target frame in the scene detection map. Through the abnormal article detection method provided by the embodiment of the invention, the abnormal article which should not appear in the scene originally can be detected, manpower is greatly liberated, and the life quality of people is improved.

The invention relates to a rock debris type identification method and apparatus, an electronic device and a storage medium. The method comprises the steps of obtaining a to-be-identified rock debris image; inputting the to-be-identified rock debris image into a pre-trained rock debris type identification model to obtain an identification result of the to-be-identified rock debris image; wherein the rock debris type identification model is obtained by training based on a ResNet34 classification network model. By means of the method, the property information of the rock debris and the rock sample of the drilled stratum can be rapidly analyzed, an effective stratum geological section is established, the defect that geological logging interpretation must depend on a deep geotechnician is overcome, and the automation and intelligence level of a drilling and logging site is improved.

The invention discloses an MRI glioma segmentation method and system based on a sparse Bayesian model and multi-map fusion, and the method comprises the steps: receiving MRI, carrying out the preprocessing, and obtaining an image which is jointly registered to a unified anatomy template through all modes of the MRI and is provided with an auxiliary diagnosis segmentation label; segmenting different modes of the MRI by adopting a sparse Bayesian model to obtain different mode segmentation result tags; and fusing segmentation results of different modes in MRI by adopting a multi-map fusion method to obtain a final glioma segmentation result with a clear boundary contour.

The present invention relates to a channel detection method and device in a system having a large number of antennas. In an embodiment, a method used in a base station having N transmitting antennas in a multiple input multiple output (MIMO) system is provided, which includes: determining NBF orthogonal pilot patterns and other M orthogonal pilot patterns; performing beamforming on NBF subspaces, and sending a pilot signal on each beam, where pilot signals on the beams of the NBF subspaces are sent respectively over the NBF pilot patterns; and sending pilot signals over at least a part of the M pilot patterns through each of the N transmitting antennas, in a manner that sub-directions in each subspace use the orthogonal pilot patterns, where a value of NBF is 10%~15% of N, and a value of M is 15%~50% of N. By using channel sparsity of the system having a large number of antennas, pilot overhead is reduced, and at least equivalent system performance can be achieved.