82results about How to "Good divisibility" patented technology

The invention discloses a vehicle detection method based on a convolutional neural network. The method includes the step S1 of collecting vehicle samples and non-vehicle samples and classifying the vehicle samples, the step S2 of preprocessing the samples, the step S3 of training a CNN vehicle detector, the step S4 calculating an average similarity table of a characteristic pattern, the step S5 of constructing a similarity characteristic pattern set, the step S6 of obtaining a CNN-OP vehicle detector, the step S7 of obtaining detection images, the step S8 of preprocessing the obtained detection images, the step S9 of constructing an image pyramid for the detection images, the step S10 of extracting characteristics, the step S11 of scanning characteristic patterns, the step S12 of classifying the characteristics, and the step S13 of combining detection windows and conducting output. An offline optimization scheme is put forward, the convolutional neural network which is completely trained is optimized, the strategy of scanning the windows after extracting the characteristics is adopted at the detection stage, and therefore the characteristics are prevented from being repeatedly calculated, and the detection speed of the system is increased.

The invention provides a liver tumor segmentation method and device based on a CT (Computed Tomography) image. The method comprises the following steps: performing Gaussian denoising on CT image data of a liver, converting the denoised CT image data into standardized data of which a gray average is 0 and a variance is 1, and performing down-sampling operation; extracting a lesion slice and a normal tissue slice from a gold standard image of the CT image of the liver, and classifying the lesion slice and the normal tissue slice into a positive sample and a negative sample; constructing a multi-level depth convolutional neural network, training a model through a stochastic gradient descent to obtain a network model, and acquiring a coarse segmentation binary image of a tumor and a pixel-classification probability image through a classifier; performing morphological erosion operation on the coarse segmentation binary image of the tumor to obtain a foreground image needed by graph cut, performing subtraction operation on the binary image of a liver and the coarse segmentation binary image of the tumor, and performing the morphological erosion operation to obtain a background image corresponding to normal tissues of the liver; and constructing an undirected graph, and obtaining a finial segmentation region of the tumor through a graph cut optimization algorithm.

The invention discloses a intra-pulse modulation and recognition method of a low signal-to-noise radar radiation source signal, which comprises the following steps of: receiving a radar radiation source pulse signal by an electronic reconnaissance receiver, after dimension reduction and A/D sampling process from radio frequency to intermediate frequency, obtaining a radar radiation source signal S(t) having different intra-pulse modulation modes, and processing the signal S(t) in a signal processing module, identifying and outputting intra-pulse modulation mode of the radar radiation source signal. With the method, the intra-pulse modulation mode of multiple radar radiation source signals can be correctly identified when the signal-to-noise is as low as -6dB; and compared with the conventional method for identifying multiple radar radiation source signals, the computational complex O (n2) or (n3) is lower.

The invention discloses a method for identifying a sound fault based on a mel energy spectrum and a convolution neural network. The method comprises the following steps: first, performing pre-emphasison initially input audio data; then, performing framing and windowing processing on the data; after that, performing fast Fourier transform on framed and windowed data; extracting energy features ona frequency domain, processing an energy spectrum by using a set of mel-scale triangular filter banks; and after that, converting the data into a Mel energy spectrum by using energy in different frequency domains corresponding to each frame as a Y axis, and different frames in a time domain as an X axis. After that, the energy spectrum needs to be further framed to adapt to an input of a CNN (convolution neural network). Each frame is a sample, and one-hot coding of a label corresponding to each sample is used as an output of the CNN to train a CNN model until a network training error reachesthe minimum. During prediction, a probability value of each type of label is output, and a label with the largest probability value is taken as a final discrimination result.

The invention discloses a radar HRRP target recognition method based on a dpLVSVM model. the method includes the steps of firstly, conducting feature extraction on radar HRRP data to obtain a power spectrum feather set X; secondly, constructing the dpLVSVM model, and obtaining the probability density function of power spectrum features and the combined condition posterior distribution of all parameters; thirdly, conducting derivation to obtain the condition posterior distribution of each parameter; fourthly, conducting circulating sampling on each parameter I times; fifthly, storing the sampling result of the parameter required by the T0th test stage; sixthly, judging whether the radar HRRP is an outside-library sample or not, if yes, rejecting the judgment, and if not, executing the seventh step; seventhly, conducting sampling to obtain the cluster mark number of the power spectrum features (please see the specification); eighthly, outputting the target classification mark number (please see the specification) of the radar HRRP. The method has the advantages of being low in classifier design complexity, good in recognition performance and good in judgment rejection performance, and can be used for radar target recognition.

The invention belongs to the technical field of detection, in particular to a method for real-time detection of weld seam targets. The method comprises the following steps: building a training sampleset: welding seam images of different shapes are collected as source samples and the source samples are pretreated to form a training sample; training detector offline: the neural network is trained under different initial conditions by using the training samples, and the optimal neural network model is obtained by multiple training as a welding seam detector; performing on-line detection: a detection image is acquired, a weld detection is performed by using the weld detector and a detection result is output. The invention overcomes the problem of poor robustness of positioning the weld seam through the morphological method, realizes the accurate positioning of the weld seam of different shapes, classifies the weld seam of different kinds accurately, and has high positioning precision by adopting the weld seam detector obtained through the weld seam image training based on the line laser imaging for the different weld seams.

The invention discloses a remote sensing image variation detecting method on the basis of weighted Gabor wavelet characteristics and two-stage clusters. The processed objects include optical remote sensing images and SAR (synthetic aperture radar) images and the remote sensing image variation detecting method includes (1) generating difference images according to remote sensing image types; (2) subjecting the difference images to Gabor wavelet transform; (3) extracting multiscale and multidirectional characteristics of the Gabor wavelet transform of the difference images; (4) designing the weighting coefficient and acquiring the weighted Gabor wavelet characteristics; (5) clustering the weight Gabor wavelet characteristics by means of the two-stage cluster strategy; (6) acquiring variation detecting results. By the remote sensing image variation detecting method, loss of marginal information is reduced, stronger, weak and slight varying areas can be detected at the same time, the total mistake pixel number is decreased, more detail information is reserved and variation results can be effectively extracted.

The invention discloses a high-resolution range profile target recognition method for kernel adaptive mean value discriminant analysis. The method comprises the steps that an original HRRP signal training set is acquired, l<2> norm normalization is performed to extract power spectrum features, and a feature sample set after preprocessing is obtained; a kernel function is adopted to perform mappingto a high-dimensional feature space; an adaptive dispersion matrix is configured; an optimal projection direction is solved; a new non-linear dimension reduction training feature set is obtained; anSVM classifier is trained; and SVM classification recognition is performed on a to-be-tested original HRRP signal. Through the method, global information of training samples is utilized in a kernel mapping space, local information is adaptively fused during information extraction, low-dimensional features with higher separability compared with common feature extraction and data dimension reductionmethods can be obtained, and recognition precision is improved. The method is also suitable for feature extraction and classification of other signals, such as classification of crack types and sizesthrough a magnetic flux leakage signal in nondestructive testing and audio signal classification.

The invention relates to a radar signal feature extraction method based on residual depth learning. With a designed residual depth learning network, deep-level feature extraction is carried out on a radar radiation source signal in a complex electromagnetic environment. The method is implemented as follows: parameters of the residual depth network are trained by using the existing radar data in adatabase; the intercepted data are sent to the input terminal of the residual depth network and results are outputted after mapping of multiple hidden layers, wherein the outputted results are used asthe depth features of the pulse string; with a clustering method, the obtained depth features are clustered, and the correlation degree of each two radiation sources after clustering is calculated and processed according to a correlation criterion; and then the parameter of each radiation source after fusion is calculated and sorting is completed. Therefore, the depth features among data can be dug; the sorting precision is high; and the radar signal feature extraction method can be applied to fields of target reconnaissance and interference source localization.

The invention relates to a battery management system for a battery having a plurality of battery cells connected in series, comprising a battery control unit, a plurality of low-voltage measuring devices, by means of which the voltage of one or more battery cells can be measured, and comprising one or more high-voltage measuring devices for a voltage across a plurality or all the battery cells, and/or at least one current measuring device, by means of which a current from or through the battery can be measured, and/or at least one a measuring module comprising a high-voltage measuring device, and a current measuring device, characterized by a signal transmission device, by way of which signals from low-voltage measuring devices and from at least one of the high-voltage measuring devices and/or the current measuring device and/or the measuring module can be transmitted to the battery control unit.

The invention relates to a combustion gas index automatic identification method based on images and belongs to the technical field of image identification. The method is characterized in that a conventional scheme in a dial positioning and character identification process is improved, the robustness of combustion gas meter image noise is improved, and a more accurate character sub-image which does not contain other impurities can be obtained; in addition, the structure information of images are utilized to describe characteristics of different character images, a classifier is constructed by a 3-neighbour method, the information separability of 10 extracted character characteristics from 0-9 is good, and the combustion gas index can be accurately extracted without a complex identification process, so that the method is much better than an existing combustion gas index identification method in efficiency, and the whole process from image collection to successful characteristic identification is no longer than 0.08s, and the combustion gas index can be integrated to a combustion gas index identification system having real-time requirements; and the method does not depend on any third party software and can be transplanted to any platform in demand such as a PC platform or an intelligent mobile phone client.

The invention relates to the fields of pattern recognition, machine learning and computer vision, and provides a zero sample classification method based on semantic enhancement of encyclopedic knowledge, which aims to solve the problem that the existing zero sample image classification method cannot give consideration to both word vector language information range and processing efficiency. The zero sample classification method comprises the steps of: S1, classifying unknown-class images by means of a trained convolutional neural network classifier, and subjecting semantic features of classification result tags to convex combination to serve as semantic features of the unknown-class images; S2, and classifying the semantic features of the unknown-class images obtained in the step S1 and semantic features in a pre-constructed semantic feature space by means of a nearest neighbor classifier, so as to obtain final classes of the unknown-class images. The zero sample classification method and a zero sample classification device provided by the invention enhance the global information of word vectors, so as to improve the accuracy of image zero sample classification.

The invention relates to a mechanical vibration signal feature extraction method based on combination of stochastic resonance and kernel principal component analysis. The method comprises the steps that firstly, the stochastic resonance method is applied for conducting pretreatment on rotor oscillation original signals measured by a sensor, the signal periodicity is improved, and the oscillation signal to noise ratio is improved; then, a time domain feature set is extracted for the pretreated output signals; then the kernel principal component analysis method is adopted for conducting nonlinear feature transformation for the extracted time domain feature set, and therefore the final needed feature set is obtained. The method is applied to feature extraction and failure diagnosis of simulated failure of an engine rotor, the result shows that the feature set extracted through the method is of linear independence, the number of dimensions is smaller, the separability is higher, the precision and efficiency of the failure diagnosis can be effectively improved, and application in engineering practice is facilitated.

The invention discloses an adaptive spectral focusing band selection method for a hyperspectral image. The method comprises the steps of firstly processing whole spectral domain hyperspectral originaldata of a ground object by utilizing three band selection algorithms, arranging results of the algorithms according to a sequence from large to small, and selecting out first N bands as optimal bandcombinations; and secondly assessing the optimal band combinations selected out by the three band selection algorithms by utilizing an independent component analysis (ICA) band assessment function, setting a J threshold, calculating out assessment results of the band combinations, and selecting the band combination with the maximum result as the optimal band combination. By utilizing an adaptive spectral focusing technology, automatic wavelength selection of wave bands can be realized during imaging of a spectrograph; a tunable light filtering component in an imaging control system is adaptively tuned to multiple spectral channels most favorable for detection and identification; spectral adaptive detection is realized; redundant and disorderly spectral information is deleted; the spectralimaging information utilization efficiency is improved; and the resource demand of information processing is reduced.

The invention discloses an acoustic target radiation noise classification method and system based on EMD (Empirical Mode Decomposition) and compressed sensing. The method comprises the following stepsof firstly, carrying out EMD decomposition on an obtained ship radiation noise signal and carrying out line spectrum component extraction on a basic mode component by utilizing compressed sensing; and utilizing the maximum mutual information coefficient to select the line spectrum component obtained by the maximum correlation coefficient of the basic mode component and the original signal to resynthesize the signal line spectrum. According to the method, the line spectrum with relatively low amplitude in the signal can be extracted; sparse dictionary training being carried out on the extracted line spectrums, sparse coding matrixes obtained after sparse representation is carried out on different signals being distributed at different positions, subsequent classification being more facilitated, time domain, frequency domain and sparse domain feature extraction being carried out on the signal line spectrums to form an accurate feature set capable of serving as a classification basis, and classification accuracy being improved; the line spectrum of the signal can be extracted more accurately; the extracted sparse features have better separability, and have a good application prospectin identification of radiation noise classification of underwater acoustic targets such as ships and warships.

The invention discloses a remote-sensing image under-segmentation object recognition method. The method comprises the following steps: data dimension reduction is performed on an image to obtain a segmentation object; clustering is performed on the segmentation object; a mixing degree index of the segmentation object is calculated according to the clustering; and the under-segmentation object is recognized in the segmentation object according to the mixing degree index. According to the judgment method of the invention, space texture and spectrum dimension information is combined, the obtained image segmentation result is more aligned with the distribution of real objects, and the under-segmentation object recognition speed and the recognition result accuracy can be improved.