49results about How to "Solve the small sample problem" patented technology

The invention belongs to the technical field of image processing and particularly provides a semantic propagation and mixed multi-instance learning-based Web image retrieval method. Web image retrieval is performed by combining visual characteristics of images with text information. The method comprises the steps of representing the images as BoW models first, then clustering the images according to visual similarity and text similarity, and propagating semantic characteristics of the images into visual eigenvectors of the images through universal visual vocabularies in a text class; and in a related feedback stage, introducing a mixed multi-instance learning algorithm, thereby solving the small sample problem in an actual retrieval process. Compared with a conventional CBIR (Content Based Image Retrieval) frame, the retrieval method has the advantages that the semantic characteristics of the images are propagated to the visual characteristics by utilizing the text information of the internet images in a cross-modal mode, and semi-supervised learning is introduced in related feedback based on multi-instance learning to cope with the small sample problem, so that a semantic gap can be effectively reduced and the Web image retrieval performance can be improved.

The invention discloses an iteration text clustering method based on self-adaptation subspace study. The method includes the following steps: (1) initiation: text linguistic data is expressed as a text vector space, initial K clusters are generated through an affine propagation clustering method, and all text clustering categories are expressed as an initial category affiliation indication matrix; and (2) iteration between the subspace projection and the clusters: the initial category affiliation indication matrix is used as prior knowledge, a maximum average neighborhood edge is used as a target to solve a subspace projection matrix, the text vector space is projected to a subspace, K clusters are generated through the affine propagation clustering method in the subspace, and a category affiliation indication matrix is updated; and a convergent function is calculated based on the subspace projection matrix and the category affiliation indication matrix till the function is converged, iteration exits, and text clustering is finished. The iteration text clustering method does not limit the capacity and distribution of text data, subspace solution and clusters are fused under a uniform frame, and an overall optimal clustering result is obtained through an iteration strategy.

The invention is suitable for the technical field of environmental monitoring, and provides an air quality PM2.5 predicating method and an air quality PM2.5 predicating system. The air quality PM2.5 predicating method comprises the following steps: performing meshing on a researched region; obtaining influence factor data in each mesh and air quality data monitored by a monitoring station; integrating the influence factor data to form a fourth-order tensor, and performing normalization processing on the four-order tensor; based on the influence factor tensor data and the PM 2.5 concentration data subjected to the normalization processing, establishing a tensor sample database; constructing an initial support tensor regression model, and training the initial support tensor regression model by taking the PM2.5 influence data tensor data and the PM2.5 concentration data as input data; determining parameters of the initial support tensor regression model through an alternating projection algorithm, and obtaining the final support tensor regression model; predicating a target sample through the final support tensor regression model to obtain the PM2.5 concentration data in the target sample.

The invention discloses a fault diagnosis method based on compressed sensing and an improved multi-scale network, which adopts compressed sensing to compress a sample, can retain interested information from an original sample signal, and improves the data analysis efficiency. A compressed sensing Gaussian measurement matrix is multiplied by a signal matrix. Due to the fact that the Gaussian measurement matrix is a random matrix, matrix elements generated each time are different, generated training samples are different, enough training samples are generated, the problem of small samples existing in the training process of the depth model is further solved, and the accuracy rate of small sample fault diagnosis is increased. Constant mapping is introduced between feature mapping with the same size and the same feature mapping number. Through feature extraction of different layers, features of different layers are connected into a whole to achieve feature fusion, a deep fusion feature vector corresponding to an input image is obtained, and therefore the problem that the multi-scale network optimization process is difficult is solved, and the fault type can be effectively recognized under different working conditions.

The invention provides a multi-linear large space feature extraction method. The method comprises the following steps: firstly, extracting gait sequence data of one period of a video stream sample, and expressing the gait sequence data as a tensor structure; secondly, preprocessing tensor data through multi-linear principal component analysis (MPCA), projecting the high-dimensional tensor data to a low-dimensional tensor structure, and removing redundancy and noise information; thirdly, optimizing the difference between inter-class Laplacian divergence and intra-class Laplacian divergence, maximizing the value of the difference, continuously performing iteration evaluation to remove convergence and achieve an iteration end condition so as to obtain projection matrixes under various modes, and obtaining a low-dimensional tensor with lower dimensionality and supervisory information through tensorial multiplication; and finally, classifying through a nearest neighbor classifier based on Euclidean distance. The method provided by the invention has a higher recognition rate compared with the MPCA, and the final feature with reduced dimensionality is shorter.

The invention discloses a communication radiation source individual identification method and system. According to the technical scheme, the method comprises the steps of: receiving communication radiation source signal, carrying out radio frequency preselection amplification, carrying out frequency mixing, carrying out intermediate frequency filtering, carrying out A/D conversion, carrying out digital orthogonal demodulation, carrying out rectangular integration double-spectrum transformation, dividing rectangular integration double-spectrum characteristic vectors into a training sample set and a testing sample set, constructing a rectangular integration double-spectrum characteristic dictionary, carrying out non-linear transformation, carrying out mapping to a cooperation projection matrix, constructing a classifier, obtaining classification residual errors, and using the type corresponding to the smallest classification residual error as type of a communication radiation source individual. According to the invention, a small sample problem in a communication radiation source individual identification process is solved, the time complexity of the algorithm is lowered, and the phase and amplitude information distortion problem in the process when an existing method based on the time domain, frequency domain or time frequency domain is used to process non-stable or non-Gaussian signals is adopted.

The invention relates to the field of biological feature identification, in particular to a face recognition method based on separation degree difference supervised locality preserving projection (SLPP). The method comprises the following steps of reading face images from face databases, conducting feature extraction on the face images to form face features, conducting the feature extraction on face region images to obtain a transformation matrix required by the feature extraction and features of training face images, conducting the feature extraction on testing face images, and classifying and recognizing through a nearest neighbor classifier based on an Euclidean distance. The method solves the problem of a small sample in face recognition, and allows an SLPP method not to be restricted by the feature dimension reserved in a PCA (Principal Component Analysis) process. The method solves the problems that the small sample results in a singular intra-class separation degree matrix, and the optimal matching dimension of PCA and the SLPP is difficult to select.

The invention provides a face recognition method based on a mutual information parameter-free locality preserving projection algorithm. According to the method, similarities among samples are calculated through adoption of MI; the similarities are directly taken as similarity coefficients among the samples; moreover, an average similarity of the samples is taken as a demarcation point, the samples are divided into neighbor samples and non-neighbor samples. A locality neighbor similarity divergence matrix and a non-neighbor divergence matrix of the samples are determined based on the neighbor samples and non-neighbor samples. According to the method, the functions of the neighbor samples are taken into consideration, and moreover, the functions of the non-neighbor samples are taken into consideration. Through application of the target function of the method, the neighbor relationships of the original neighbor samples are kept after projection; and the non-neighbor samples are kept away as far as possible after projection. With respect to solution of the target function, the dimensions of the samples are reduced to a non-zero space of an overhaul divergence matrix by employing a PCA algorithm, and then the target function is transformed into a difference form. The small sample problem is effectively solved. No any parameter needs to be set, and the practicability of the method is enhanced.

The invention discloses a reliability evaluation method for structural mechanism products of carrier rockets to achieve the purpose of quantitative evaluation of the reliability of various mechanism products of carrier rockets. The method comprises the following steps that firstly, the three-element information of products is determined, and a failure mode distribution type, performance indexes and types which a functional type belongs to of the products are listed; secondly, based on the priority selection principle, applicable evaluation models are determined, wherein the corresponding evaluation models are determined in the sequence of A, the failure mode distribution type, B, the performance indexes and C, the functional type; thirdly, needed test information corresponding to the reliability models is collected; fourthly, reliability calculation is carried out according to the selected reliability models and reliability data. By means of the method, the current situation that at present, qualitative evaluation of structural mechanism products of carrier rocket serves as the main part, and reliability evaluation is insufficient in systematicness and normalization is changed, and the characteristic of small samples of mechanism products of carrier rockets can be met.

The invention discloses an improved compressed sensing-based face recognition method. The method includes the following steps that: step 1, image pre processing is carried out: a cumulative distribution function is adopted to perform histogram equalization on a gray face image, and then, gray face image feature vectors extracted by an invariant uniform pattern are rotated through adopting an LBP (local binary pattern) so as to construct an LBP feature space; step 2, neighbor samples of a test image are searched in the low-dimensional LBP feature space; and step 3, the neighboring samples are utilized to adaptively construct a redundant dictionary, and the sensing recognition of the test image is completed. According to the method, the neighbor samples of the test sample are searched in the low-dimensional LBP feature space, and are adopted to form the compete redundant dictionary, and therefore, the number of the atoms of the dictionary is greatly decreased, and at the same time, the atoms of the dictionary have higher structural similarity with the test sample, and thus, the algorithm can not only improve recognition speed, but also can improve correct recognition rate.

The invention provides a fault diagnosis method for a power compartment of an armored car of a related vector machine based on optimization, and the method comprises the steps: 1, collecting the fault sample data of the power compartment through a sensor, and obtaining sample data S={(xi, yi)}, wherein xi is an i-th n-dimensional attribute sample, i is within the range [1, N], N is the total number of samples, and yi is a fault type corresponding to the i-th sample; 2, carrying out the normalization preprocessing of the sample data S, and obtaining a training set and a test set; 3, selecting an optimal RVM core parameter sigma through a cuckoo search algorithm based on Gaussian disturbance; 4, inputting the data of the training set into an RVM model, carrying out training, and constructing the related vector machine; 5, carrying out the classification of the data of the test set through employing the constructed RVM (related vector machine), and obtaining a corresponding fault state of the power compartment. The method can improve the classification precision while shortening the training time, is strong in generalization capability, can accurately detect different faults of the power compartment, and well solves a problem of fault diagnosis of an integrated machine of the power compartment.

The invention discloses a facial expression recognition method based on MB-2DPCA features. The method comprises the following steps: carrying out local texture feature extraction on a facial expression image through a multi-scale parameter MB-LBP operator; carrying out feature extraction on expression features and constructing an MB-LBP feature space of the image; calculating a projection matrix W through a 2DPCA algorithm in training samples, and mapping all of the training samples to a subspace through the projection matrix W, and obtaining feature expression of the expression image in the subspace; and mapping a test sample to the subspace through the projection matrix W, and carrying out classification on the features of the test sample through a sparse representation classifier to obtain the category of the test sample. The method prevents the problem of large covariance matrix calculation dimensions due to conversion from an image matrix to the image vector, and the problem of small samples due to fewer samples; and the calculation time is less.

The invention discloses an abstract image emotion recognition method, comprising the following steps: a convolution neural network is pre-trained by using a natural image data set; style features areextracted from the natural image emotion recognition data set and the abstract image emotion recognition data set, and the style differences between each sample in the abstract image emotion recognition data set and each sample in the natural image emotion recognition data set are calculated; a subset of a natural image emotion recognition data set most similar to a style characteristic of that abstract image emotion recognition data set is selected, and the natural image emotion recognition data set subset and the abstract image emotion recognition data set are used for two-layer migration and joint fine tune to obtain an abstract image emotion recognition model. Compared with the prior art, the invention improves the identification accuracy.

The invention relates to a face recognition method that suits for the small sample, which uses two-dimensional matrix, to establish sample data matrix, similar member matrix, getting typical correlation matrix and finally generalized inverse matrix. It solves the problem that the queer covariance matrix is added artificial disturbance to process correction, which brings disturbance data introducing more random and uncertainty, making the result of small sample is not satisfied, leading to a significant increase in calculation, discarding the image null space ahead due to the method of pre-treating samples by dimensionality reduction, losing some important distinguishing information. The invention replaces number 1 with a matrix, that is the same size as sample image, to make the similar member matrix not only reflect the affiliation on each sample and any kind, but retain the space information of rows and columns in the sample images.

The invention provides an image data classification system based on Universum with combination of a matrix Ho-Kashyap algorithm. Firstly, a certain number of a third kind of sample points located between two kinds of samples, namely Universum samples, are generated by employing an In-Between-based generation strategy; the Universum sample points are substituted into a regularization item Runi; the regularization item is introduced into a HK classification model after matrixing to form a complete matrixing HK model with combination of Universum; and finally, the model is trained, optimal parameters of the model for a current training data set are obtained, and an optimal classification decision surface is generated. In the test stage, the test sample points are substituted into a decision surface function for determination, and classification labels are output. Compared with the conventional classification technology, according to the system, the Universum samples are introduced, the comparison of the original two kinds of samples is more obvious, and the accuracy is further improved.

The invention belongs to the field of natural language processing, and particularly relates to an entity recognition method for naming liver cancer pathological texts, which comprises the following steps: acquiring liver cancer pathology text information in real time, and preprocessing the text information; inputting the preprocessed text information into the trained liver cancer pathology text named entity model to obtain a liver cancer pathology text information recognition result; classifying and marking the liver cancer pathology text information according to an identification result. According to the invention, ten types of key entities of the liver cancer pathological text can be accurately recognized and extracted, and the entity recognition effect is good.

The invention discloses an image sample generation method and device based on a partial shadow special effect, computer equipment and a storage medium, and relates to an artificial intelligence technology, and the method comprises the steps: calling a contrast brightness adjustment algorithm, carrying out contrast brightness adjustment of an input image, and obtaining a corresponding dark image and a corresponding bright image; obtaining an original picture size of the input picture, and initializing according to the original picture size to obtain an initial fuzzy forward dodging picture so as to perform fuzzy forward dodging on the initial fuzzy forward dodging picture to obtain a current fuzzy forward dodging picture; performing Gaussian blur on the current fuzzy forward dodging pictureto obtain a Gaussian blur picture; and synthesizing according to the dark picture and the bright picture corresponding to the input picture and the Gaussian blur picture to obtain an image sample corresponding to the input picture. According to the invention, expansion of the image sample of the input picture based on the local shadow special effect is realized, the acquisition difficulty of acquiring the identity card picture sample is reduced, the expanded image sample is close to a sample shot in a real scene, and the problem of small samples is solved.