97results about How to "Reduce the number of features" patented technology

The invention discloses a hyperspectral image super-resolution restoration method based on a 3D convolutional neural network. According to the technical scheme, the 3D residual dense network is characterized by comprising 3D convolution kernel to convolve the hyperspectral image spectral dimension and 3D sub-pixel recombination to enlarge the image and reconstruct the high-resolution image part, and unifying the two parts in the deep convolutional neural network framework 3D-RDN; hierarchical characteristics of the convolutional layer are fully utilized through structures such as residual dense blocks, and super-resolution restoration of the hyperspectral image is achieved. At present, when an existing method based on deep learning is applied to a hyperspectral image, the characteristics of the hyperspectral image are not fully considered, and therefore it is difficult to effectively utilize rich spectral dimension information of the hyperspectral image to reconstruct a high-resolutionimage. According to the method, all spatial spectrum information of the hyperspectral image is fully utilized, efficient super-resolution restoration is achieved, and the PSNR value is superior to that of an existing method.

The invention belongs to the technical field of mode recognition and information processing and provides a mass image retrieval system based on cluster compactness. Steps include 1, calculating local features of images in a sample image library and a test image library; 2, calculating cluster compactness of each image, namely clustering the local features to acquire each type of cluster centers, counting a local feature distribution histogram and spatial statistical information of each cluster, and generating cluster compactness; 3, randomly sampling cluster compactness of the sample image library, clustering components of the cluster centers in the sampled cluster compactness to generate a vocabulary tree, and quantizing the cluster compactness of the images in the test image library to the vocabulary tree to generate corresponding inverted files; 4, retrieving by a modified retrieval algorithm based on the vocabulary tree, namely retrieving, by retrieving the inverted files in the vocabulary tree and calculating the weight of similarity between retrieval images and the image library image cluster compactness.

The invention discloses a topic model-fused scene image classification method, and relates to the field of deep learning and image classification. The method comprises the following steps of: preprocessing data sets, and expanding the quantity of obtained data sets so as to obtain an image data format according with deep learning model processing; constructing a convolutional neural network modelaccording with scene image classification, and pre-training the processed image data sets by using a convolutional neural network; and carrying out end-to-end iterative training on the constructed convolutional neural network by using training sets, adjusting parameters in the network, verifying the trained model by using verification sets, modeling extracted scene image features with discrimination ability, extracting hidden topic variables between the features and images so as to obtain image topic distribution represented by a k-dimensional vector, wherein k represents a topic quantity, each image can be considered as a probability distribution diagram formed by a plurality of topics, and scene image classification is realized by utilizing a classifier.

The invention discloses an intelligent rotary machine fault depth network feature identification method. A vibration sensor is arranged at a to-be-detected rotating mechanical part of a train rollingbearing; collecting an original vibration sequence when the rolling bearing works; decomposing and reconstructing the original vibration sequence through a singular spectrum analysis method; extracting a root-mean-square value of the reconstructed vibration sequence; standard deviation, skewness index and peak value; a fault position is judged by using a rotary machine fault position diagnosis model obtained by training of a support vector machine; and then, carrying out ensemble empirical mode decomposition on the reconstructed vibration sequence, calculating the permutation entropy values ofa group of decomposed intrinsic mode components, taking the permutation combination of the permutation entropy values as a detection characteristic, and judging the fault type by using a rotary machine fault type diagnosis model obtained by training of a support vector machine. The fault position and the fault type of the rotary machine can be detected more timely, and the fault diagnosis accuracy and reliability are improved.

The invention discloses a face recognition model construction method and device, a face recognition method and device and a storage medium. The method comprises steps of preprocessing face pictures, extracting four local areas, namely a left eye area, a right eye area, a nose area and a mouth area, of the preprocessed face image to perform face partitioning; amplifying the data on the basis of a partitioning result, respectively amplifying the data by considering the shielding of 1/2 and 1/4 conditions, and finally training a deep learning neural network by using the amplified data to construct a training feature data set, thereby realizing the face recognition based on feature matching. The method is good in shielded face recognition effect, does not need a large number of shielded samples, and is small in occupied memory.

The invention relates to an image recognition method and device. The method comprises the following steps that: the features of a target image are extracted; dimensionality reduction is performed on the features of the target image according to a principal component analysis method; and the target image which has been subjected to dimensionality reduction is recognized according to a pre-trained image model. With the image recognition method adopted, a terminal can perform dimensionality reduction on the features of the target image according to the principal component analysis method, and therefore, the number of the features of the target image can be decreased, and the key features of the target image can be reserved; the terminal can recognize the target image which has been subjected to dimensionality reduction according to the pre-trained image model; and therefore, the efficiency of image recognition can be improved, and the accuracy of image recognition will not be reduced, and the use experience of a user can be improved.

The invention discloses a playback speech detection method. At a training stage, a first variation coefficient vector, a normalized first cepstrum feature matrix, a second variation coefficient vectorand a normalized second cepstrum feature matrix of each speech sample in a speech database are firstly acquired as four kinds of features; the four kinds of features of all positive samples are theninputted to a GMM model for training, four positive sample feature models are obtained, and besides, four negative sample feature models are obtained; at a testing stage, four kinds of features for to-be-detected speech are acquired in the same way, the kinds of features are inputted to the corresponding positive sample feature models and the negative sample feature models respectively, and four likelihood ratio scores are obtained; and according to the four likelihood ratio scores, a final score is obtained, and through compairing the final score and a judgment threshold, whether to be a playback speech is judged. The playback speech detection method is not only limited to a text-related voiceprint authentication system, and has the advantages of low equal error detection possibility, strong robustness and relatively low calculation complexity.

The invention provides a disk fault prediction method and system, a terminal and a storage medium. The disk fault prediction method comprises the steps of screening SMART feature items related to diskfault prediction by utilizing a recursive feature elimination algorithm; generating a disk health parameter according to the fluctuation condition of the related feature item of the disk in the preset time period; training a neural network model by using the historical disk SMART feature data; and inputting the disk health parameters into a neural network model to obtain a disk fault prediction result. The SMART features are screened by using the RFE algorithm, the number of the features is reduced, the calculated amount is reduced, overfitting is inhibited, the new feature value is generatedaccording to the change times of the SMART feature value within seven days, and the proportion, the balance accuracy and the recall rate of two types of errors in the loss function are trained and adjusted more easily.

The invention relates to a power system reliability evaluation technology, in particular to a power system risk evaluation method based on cross entropy theory optimization support vector machine, wherein the cross entropy method CEM is applied to a support vector machine SVM to carry out feature selection and SVM parameter optimization; the n-dimensional feature quantity is used as the input of SVM, and the training model is generated by off-line training. The risk level of the power system under the current state is predicted by learning the historical samples. The method includes generatingoff-line data, training the risk sample data by using cross-entropy optimization SVM method, and obtaining the optimized characteristics and parameters of power network risk assessment model; the on-line real-time power system risk assessment is carried out. Using cross-entropy algorithm to optimize support vector machine for risk prediction can effectively remove redundant and irrelevant features, reduce the number of features, combine the optimized parameters can have good convergence, and reduce the computational complexity and time-consuming. It has the characteristics of low error rate and short test time.

The invention discloses a driver eye state monitoring method based on invariant moment. After an facial image is preprocessed, an eye area is extracted out according to an image threshold segmentation method based on the two-dimension digital image fractional order integral and the Legendre moment, first three orders of central moments and four feature quantities of the eye area are combined to serve as a feature vector for matching recognition of eye states, Euclidean distances between feature vectors of an area to be detected and feature vectors of a template image are calculated one by one, and thus the eye state of a driver is judged out. The respective advantages of the central moments and the feature quantities of the eye area are combined, the number of features for matching recognition is decreased, and space dimensionality of the features is reduced; by introducing the Euclidean distances between the feature vectors of the candidate eye area and the feature vectors of the eye template image, algorithm complexity is further lowered, and the system recognition speed is increased.

The invention discloses a text classification method based on an improved firefly algorithm and K neighbors. A text feature selection model is constructed by combining information gain and the fireflyalgorithm. The method comprises the following steps: all features are sorted by using information gain, and then a more representative feature subset on a feature set sorted in the front is found out by using the relatively strong optimization capability of an improved firefly algorithm. The step length factor alpha in the firefly algorithm is adjusted, so that the global search capability of the algorithm is ensured, and the local search capability is also ensured. A new fitness function is introduced, so that the dimensionality of the features is properly reduced on the basis of improvingthe precision of the feature subsets. And finally, the model is used for text feature selection, and the obtained feature subset is used for KNN text classification. According to the method, the defects that a firefly algorithm is prone to early maturing and falling into local optimum, the convergence speed is low and the like in the process of searching for the optimal text feature subset can bewell overcome, so that a more accurate subset is obtained, and the text classification accuracy is improved.

The invention discloses an image segmentation method based on background detection. The method comprises the following steps: acquiring N silicon wafer images shot by X-ray; respectively carrying outcolor image segmentation on the N silicon wafer images, screening out a single-channel image with the maximum contrast ratio from the N silicon wafer images, and performing further graying to obtain Nfirst target images; carrying out global threshold processing on the N first target images, carrying out hole filling until no gap exists in pixels, and obtaining N second target images; constructinga connected domain of the N second target images according to a set pixel adjacency relation, and further screening out a silicon wafer region by utilizing the area characteristics of the silicon wafer to obtain N third target images; and performing morphological opening operation on the N third target images, reducing the definition domain of the N third target images to a set definition domain,and performing matting processing to obtain N target images. According to the method and device, the original image is subjected to series processing, so that the feature number is reduced on the premise of ensuring the image quality, and the target image is quickly segmented from the original image.

The invention discloses a malicious code obfuscation feature cleaning method, and belongs to the field of machine learning information safety. The method involves a feature selection method and an obfuscation feature cleaning method, and the effectiveness of a traditional malicious code feature extracting method is improved. Compared with the traditional malicious code feature extracting method, the malicious code obfuscation feature cleaning method can effectively prolong the effective time limit of a malicious code feature extracting algorithm, and improve the interference resistance of thefeature extracting algorithm. Firstly, a feature library is built through an n-gram feature extracting method. Since the feature extracting algorithm cannot solve the obfuscation operation problem ofmalicious codes, the feature library contains a large number of obfuscation feature values of the malicious codes. Through an obfuscation feature cleaning algorithm, the interference of abnormal datain a model identification rule can be removed. On this basis, from the aspect of the scale of a training dataset, a feature selection method is put forward. By means of the malicious code obfuscationfeature cleaning method, on the basis of guaranteeing that the model identification precision is not lowered, the number of features which are finally used in the model is effectively lowered.

The invention discloses a representation and classification method of high-order and high-dimensional image data, belonging to the field of pattern recognition, which solves the problems of noise influence in the process of image recognition, destruction of original image structure, high data feature dimension, incompatibility between expression algorithm and classification algorithm and the like.The technical scheme adopted by the invention is that the image data is projected from the original space into the low-rank subspace according to a pattern by using a projection matrix set to obtaina low-rank representation of the data; a sparse representation dictionary and a linear classifier are trained in the low rank subspace to classify the image data. The image data is projected from theoriginal space to the low rank subspace by using the projection matrix set, and the low rank representation of the image data is obtained. Sparse representation dictionaries and linear classifiers aretrained in low rank subspaces to classify image data.