53results about How to "Reduce feature redundancy" patented technology

The invention provides a rolling bearing fault diagnosis algorithm based on time-frequency domain multidimensional fault feature fusion. Aiming at the respective features of vibration signals of a rolling bearing in a normal state, a roller fault state, an inner ring fault state and an outer ring fault state in a time-frequency domain, through extraction of time domain and frequency domain features, redundancy removal and re-fusion, fault features are described in an optimal way to obtain an intelligent judgment result. First, wavelet de-noising is performed on extracted original rolling bearing vibration data; then, time domain feature vectors are extracted to form a time domain feature matrix, and coefficient energy moments after wavelet packet decomposition and reconstruction are extracted to form a frequency domain feature matrix; and the time and frequency domain matrixes are further fused to obtain a time-frequency domain multidimensional fault feature matrix. Redundancy of the multidimensional feature matrix is eliminated to obtain a new multidimensional feature matrix. Then, information of multidimensional features is fused with a weighted feature index distance, and a state judgment result of the rolling bearing is obtained through the feature index distance obtained through fusion.

The invention discloses a method for automatically detecting and classifying textile flaws based on pattern recognition and image processing, which is conducted according to the following steps: 1) constructing a device based on the pattern recognition and the image processing; 2) collecting RGB (red-green-blue) color fabric images with the resolution of 2048*2048; 3) preprocessing the collected color fabric images; 4) establishing a method base; 5) searching and locating flaw zones on cloth; 6) collecting the characteristic values of the flaw zones on the cloth; 7) processing the collected characteristic values; and 8) using a 'quantitative conjugate BP (back propagation) neural network algorithm' to classify the data after dimensionality reduction in the step 7). By using the method, the flaws on the cloth can be detected and classified, and the method has the advantages of high detection speed, clear classification and low manual labor intensity.

The invention discloses a retinal vessel segmentation method based on combination of deep learning and a traditional method and relates to the fields of computer vision and mode recognition. According to the method, two grayscale images are both used as training samples of a network, corresponding data amplification, including elastic deformation, smooth filtering, etc., is done against the problem of less retinal image data, and wide applicability of the method is improved. According to the method, an FCN-HNED retinal vessel segmentation deep network is constructed, an autonomous learning process is realized to a great extent through the network, convolutional features of a whole image can be shared, feature redundancy can be reduced, the category of multiple pixels can be recovered from the abstract features, a CLAHE graph and a gauss matched filtering graph of the retinal vessel image are input into the network, an obtained vessel segmentation graph is subjected to weighted average, and therefore a better and more intact retinal vessel segmentation probability graph is obtained. Through the processing mode, the robustness and accuracy of vessel segmentation are improved to a great extent.

The invention discloses a multi-modal trajectory prediction method for pedestrians in a complex scene. The method comprises the following steps: performing picture feature extraction by using a 16-layer convolutional neural network of a visual geometry group; performing feature processing on the trajectory data by using a full connection layer; inputting a trajectory data feature vector VS to enter a generative adversarial network to complete a coding and decoding network function; inputting picture feature data and track feature data to physics, wherein a social attention module considers terrain limitation and pedestrian interaction; obtaining a better track generation prediction result through the updated generator part; and obtaining a stable trajectory prediction model SPM. Accordingto the method, the prediction precision can be effectively improved, a plurality of reasonable prediction tracks can be generated, the related terrain limitation information can be extracted accordingto the feature information of the original picture, and the social interaction situation between different pedestrians in the same complex scene can be considered. The method can predict the future track of the pedestrian more quickly and accurately.

The invention discloses a method and a device for training a recommendation model, and a recommendation system. The device for training the recommendation model comprises a log acquisition unit, a first characteristic extraction unit, a second characteristic extraction unit, a sample generation unit, a sample aggregation unit and a model training unit, wherein the log acquisition unit is suitable for acquiring a plurality of access logs; the first characteristic extraction unit is suitable for extracting a first characteristic used for identifying recommendation information which presents on a corresponding page of each access log from each access log, and extracting a second characteristic used for identifying the context information of the page; and the second characteristic extraction unit is suitable for acquiring the interest degree sorting value of the user associated with each access log on various pieces of recommendation information, inquiring the interest degree sorting value corresponding to the recommendation information which presents on the corresponding page of the access log, and taking the acquired interest degree sorting value as a third characteristic.

The invention provides a multi-mode ultrasonic image classification method and a breast cancer diagnosis device, and the method comprises the steps: S1, segmenting a region-of-interest image from an original gray-scale ultrasonic-elastic imaging image pair, and obtaining a pure elastic imaging image according to the segmented region-of-interest image; s2, extracting single-mode image features of the gray-scale ultrasonic image and the elastic imaging image by using a DenseNet network; s3, constructing a resistance loss function and an orthogonality constraint function, and extracting shared features between the gray-scale ultrasonic image and the elastic imaging image; and S4, constructing a multi-task learning framework, splicing the inter-modal shared features obtained in the S3 and thesingle-modal features obtained in the S2, inputting the inter-modal shared features and the single-modal features into a plurality of classifiers together, and performing benign and malignant classification respectively. According to the method, benign and malignant classification can be carried out on the gray-scale ultrasonic image, the elastic imaging image and the two modal images at the sametime, and the method has excellent performance of high accuracy and wide application range.

The invention discloses an urban noise identification method based on deep network migration features and augmented self-encoding. The method comprises the following steps: 1, preprocessing each typeof collected urban noise signals, including denoising, framing and windowing; 2, converting the processed noise signal into a spectrogram; 3, performing feature extraction on the spectrogram obtainedin the step 2 by using a plurality of pre-trained deep convolutional neural networks; 4, fusing the obtained feature vectors x by using an augmented auto-encoder; 5, on the basis of the fusion features in the step 4, constructing a multilayer one-class classification model; 6, calculating an output weight and a decision threshold value of ML-OCRLS; and 7, carrying out classification prediction onunknown signals. The hidden layer neurons of the augmented auto-encoder provided by the invention can adjust and optimize all features, main information can be extracted based on ML-OCRLS of the augmented auto-encoder, feature redundancy is reduced, and meanwhile, multiple transfer learning features are effectively fused, so that the classification precision of a classifier is improved.

The invention relates to a high-resolution remote sensing image feature selection method based on a Cramer's V index. The method comprises the following steps of: 1) carrying out pretreatment and image feature extraction on an obtained remote sensing image; 2) carrying out continuous feature discretization parallel processing based on the Cramer's V correlation index; 3) carrying out parallel processing to obtain two contingency tables between two factures; and 4) carrying out feature selection based on the Cramer's V correlation index. The method is good in feature selection effect, high in efficiency and high in applicability, and is capable of efficiently improving the classification accuracy of the remote sensing image. Besides the remote sensing processing, the method can be widely applied to various problems such as pattern classifications of various high-dimensionality and complex type data sets (such as texts, images, medical diagnosis, and genetic data), data mining and visualizing, and the like.

The invention discloses a dynamic gesture recognition method and system based on a two-dimensional convolutional network, and the method comprises the steps: collecting an actual dynamic gesture video, and carrying out the framing processing of the video; carrying out frame sampling on the actual image after framing; encoding the actual image after frame sampling to obtain an actual feature vectorof the actual image; fusing the actual feature vectors to obtain an actual feature matrix; and inputting the actual feature matrix into the trained two-dimensional convolutional neural network, and outputting a gesture recognition result. By processing a source video stream into a frame of image, and sending the frame of image into a two-dimensional convolutional network, a classification resultof a gesture action is obtained; wherein the image generated by the video contains the spatial feature information and the time sequence information of the video. According to the method, the calculation complexity of gesture recognition is effectively reduced.

The invention discloses a large-scene point cloud semantic segmentation method. The method comprises the following steps: carrying out feature splicing on three-dimensional point cloud data containing feature information to obtain point cloud initial features; performing expansion graph convolution and random sampling on the point cloud initial features to obtain multi-layer intermediate features and sampling coding features; performing cross-layer context reasoning on the multi-layer intermediate features to obtain complementary context features, and splicing the complementary context features into the sampling coding features obtained in the last layer to obtain final coding features; decoding the final coding feature to obtain a decoding feature; inputting the decoding feature into a full connection layer classifier to obtain segmentation result prediction; and constructing a loss function, training and optimizing the model, and storing model parameters. According to the method, cross-layer context reasoning is used for aggregating multiple layers of contexts in the coding stage, attention fusion is adopted for feature selection in the decoding stage, information loss can be effectively made up and feature redundancy can be effectively reduced while efficiency is guaranteed, and therefore the accuracy rate is improved.

The invention discloses a human action recognition method based on a multi-feature motion-in-depth diagram. The method specifically includes steps of (1), projecting all the frames of a depth video tofront, side and top orthogonal planes; (2), stacking the absolute pixel differences of two continuous projection frames on each plane to form the motion-in-depth diagram (DMM{v=f, s and t}), and respectively extracting LBP characteristics, GIST characteristics and HOG characteristics to form characteristic descriptors corresponding the three directions; (3), using a relative entropy principal component analysis method to perform characteristic fusion and dimensionality reduction on the three different types of the characteristic descriptors; (4), calculating the sparse reconstruction errors of the characteristics of each type of action training samples under an l1 norm and an l2 norm after the fusion and dimensionality reduction, and using an entropy weight method to perform adaptive weighted fusion on the two types of reconstruction errors to design an error-fusion-based classifier for action recognition. By the method, the accuracy of human action recognition can be improved.

The invention relates to a fish identification method and system based on wavelet packet multi-scale information entropy. The method, based on an active sonar identification strategy, comprises step 101): sending a sound signal to an underwater fish body and obtaining an echo signal backscattered from the underwater fish body; step 102): performing a filtering and an amplitude normalization preprocessing on the collected echo data of the fish body; step 103): extracting features of the preprocessed echo signal with a specific method being as follows: performing a wavelet packet multi-scale decomposition on the preprocessed echo signal to obtain a wavelet packet coefficient, and reconstituting the wavelet packet coefficient to obtain a time-domain signal in each frequency band of the fish body echo signal, and then extracting information entropy of the time-domain signal obtained through reconstituting the wavelet packet coefficient, wherein the information entropy constitutes a feature quantity; and step 104): inputting the feature quantity into a classifier to perform a classification treatment so that kind identification of the underwater fish body is completed.

The invention belongs to the technical field of birthmark-based software recognition, discloses a constraint limited clustering and information measuring software birthmark feature selection method, and aims at measuring distances between same-class features and distances between different-class features on the basis of mutual information by adoption of constraint limited clustering analysis. Themethod comprises the following steps of: during software feature selection, carrying out equivalent semantic transformation on software; carrying out feature segmentation; combining a program slice technology to carry out limited group classification on features; constructing gain function and penalty function evaluations for segmented feature fragment sets; and carrying out different constructiongroup-based hierarchical clustering selection so as to screen invariable features in same classes and get rid of common features in different classes. According to the method, relevancy between features is considered, and in a set formed by the screened birthmark features, the differentiation information amount is the maximum and the redundancy is the minimum, so that the anti-attack performanceof the birthmark features is ensured and the uniqueness of the birthmark features is ensured. According to the method, the robustness and credibility of the software birthmark features are improved, and the feature-based software recognition rate is greatly improved.

The invention discloses a lithium ion battery RUL prediction method based on HI and ANN, and the method comprises the steps: firstly searching all measured voltages, currents, temperatures and the like in a charging and discharging process, and proposing different feature extraction methods according to different characteristics of data to obtain HI related to the capacity. For abnormal values possibly appearing in the extracted HI, a common average value substitution method and a normalization substitution method are adopted according to the positions of the abnormal values; then, based on the corrected Pearson correlation coefficient between each HI and the capacity, optimal feature selection is carried out, and the HI with the maximum value is selected as the optimal HI; and finally, a neural network is constructed from three aspects of shape features, time dependence features and sequence transformation features of the time sequence HI, so that prediction of the RUL is achieved. The extraction method used in the invention is relatively simple; the subsequent network training and prediction cost is reduced; and the interference of noise on the sequence is reduced, and the prediction precision is improved.

A supervision multi-view feature selection method based on automatic generation of view and unit with l1 and l2 norm minimization belongs to the technical field of remote sensing image data processing. The method provided by the invention is to solve a problem of information loss of a high-resolution remote sensing image in the feature selection process. The method comprises three steps: 1. acquiring remote sensing image data and pre-processing the same; then carrying out feature extraction to obtain an feature vector set; and normalizing all feature vectors of the feature vector set to obtain an original feature vector set; 2. generating multiple views from the original feature vector set obtained in step 1 by adopting an affinity propagation algorithm; and and 3. carrying out supervision multi-view feature selection on the multiple views obtained in step 2 based on the l1 and l2 norms. The invention relates to the supervision multi-view feature selection method.

The invention belongs to the field of computer vision and depth learning, the invention particularly discloses a hand-drawn sketch retrieval method based on deformable convolution and depth network, The method comprises the following steps: S1, acquiring a hand-drawn sketch and natural color map database S2, A natural color image is converted into an edge map S3 by an edge detection algorithm, Thedepth network S5 based on deformable convolution is trained by preprocessing S4 of the hand-drawn sketch and the edge map through morphological operation, and the depth features S6 of the hand-drawnsketch and the edge map of the natural image are extracted respectively by using the trained depth network, and similarity between the features is calculated and the retrieval result is returned. Themethod of the invention has the beneficial effects that deformable convolution is incorporated into the traditional neural network, the limitation of the standard convolution on the hand-drawn sketchcan be broken, the robustness of extracting features of the network hand-drawn image can be improved, and the feature redundancy can be reduced. The network structure proposed by the invention can greatly improve the retrieval precision of the hand-drawn sketch.

The invention discloses a panoramic image segmentation method, and the method comprises the steps: extracting a multi-scale feature map of an input image, and extracting a region candidate box from the multi-scale feature map; extracting a region candidate feature map from the multi-scale feature map according to a region candidate box, and segmenting a foreground image from the region candidate feature map; determining a foreground feature map and a background feature map based on the multi-scale feature map; determining a background combined feature map based on the foreground feature map and the background feature map, and segmenting a background image from the background combined feature map; and obtaining a panoramic segmentation image according to the foreground image and the background image. The high-precision panoramic segmentation image can be quickly obtained.

The invention provides a hyperspectral image classification method based on a self-paced learning double-flow multi-scale dense connection network, and the method comprises the following steps: 1, reading a hyperspectral image for preprocessing, including unsupervised waveband selection dimension reduction and image partitioning, and dividing a training set and a test set; 2, establishing a local spatial spectrum feature extraction branch and a global spatial feature extraction branch, adaptively fusing the features of the two branches, and finally connecting full connection and a Softmax layer; 3, respectively inputting the local image block and the global image block corresponding to the same pixel point into two branches, training an effective hyperspectral image classification model by taking a minimized cross entropy loss function as a target, and selecting training data by utilizing a self-paced learning framework; and 4, after the training is finished, inputting the test set data into the network to obtain a prediction category output by the model.

The invention discloses a tone mapping image quality evaluation method based on image segmentation. Aiming at the characteristic that main distortion types of different regions of a tone mapping imageare different, the tone mapping image is divided into a complex region and a flat region, texture detail features are extracted from the complex region, chrominance features are extracted from the flat region, and then the texture detail features and the chrominance features are extracted from a global region. The method aims at the characteristic that detail distortion of highlight and low-darkareas of an image is too large. An image is divided into a highlight area, a low dark area and other areas. Information entropy features are extracted from different areas to represent the distortiondegree of the image, then a high-brightness low-dark area threshold value serves as a feature to judge the brightness distribution uniformity degree of the image, feature values with good effects whendifferent areas are evaluated are reserved, feature values with poor effects are removed, and feature redundancy is reduced; and the correlation between the obtained objective evaluation result and the subjective perception of human eyes is effectively improved.

The invention discloses a patient physiological data feature selection method based on a distance greedy strategy. Aiming at the disadvantage of a conventional feature selection algorithm in performance, the invention achieves the improvement, and the algorithm employs the greedy strategy for improving a position updating part in an original grey wolf algorithm, thereby improving the capability ofthe algorithm in developing the optimal solution, and improving the convergence rate. Moreover, the algorithm can effectively improve the classification accuracy, and reduces the data feature redundancy.

The invention discloses a video target real-time tracking method and system based on depth feature fusion and adaptive correlation filtering. The system comprises a feature extraction module, a feature fusion module and a correlation filtering module. The feature extraction module uses a lightweight network model to extract multi-layer depth features, and the real-time performance of a target tracking task is ensured. The feature fusion module provides a multi-layer depth feature fusion strategy of typical correlation analysis for the problem that independently extracted multi-layer features are not complete enough in representation of a target. According to the method, the target expression capability and the target and background distinguishing capability are improved, the feature redundancy is reduced, and the calculation amount of subsequent related filters is reduced. The correlation filtering module provides a correlation filter updating strategy based on response value dispersion analysis for solving the tracker drifting problem caused by challenges that the target is deformed, shielded, moved out of the view, and is rotated existing in a target tracking task, filter template updating is carried out in a self-adaptive mode, and the specific problems are relieved.

The invention discloses a traffic sign and signal lamp detection and identification optimization algorithm. The algorithm is divided into a feature extraction stage, a region candidate stage and a hierarchical classification stage. In the feature extraction stage, a Ghost bottleneck module is introduced to construct a feature extraction network, big data training is carried out on ImageNet to obtain a pre-training model, image features are extracted through the pre-trained feature extraction network, and pooling processing is carried out on a feature map; in the region candidate stage, RPN subnets are adopted to obtain candidate regions, and feature maps corresponding to the candidate regions are cut and scaled, so that the sizes of the feature submaps to be classified are the same; in thehierarchical classification stage, images are classified through a hierarchical classification method, and traffic lights and traffic signs are identified. Compared with a baseline algorithm, the performance of the algorithm is greatly improved in all aspects, and the real-time performance and reliability requirements of an automatic driving system can be met.

The invention discloses a dual-polarization SAR small ship detection method based on an enhanced feature pyramid, and aims to solve the problem that small and medium-sized ships in a synthetic aperture radar (SAR) image are generally weak in scattering intensity and only occupy a small number of pixels in the image to cause leak detection. Two channel fusion coefficients in dual-polarization SAR data are automatically learned through feedback of a detection result, and a fusion channel input feature extraction network is obtained through coefficient weighting. In the improved attention enhancement type low-level feature pyramid, the sampling frequency under the deep network is reduced to obtain the low-level feature pyramid, prediction is carried out on a large-scale feature map, the problem that features disappear due to the fact that small ships occupy few pixels in an SAR image can be relieved, and meanwhile a space and channel attention mechanism is introduced.

The invention discloses a fish identification method and system based on segmented time-domain centroid features. The method, which is an identification strategy based on an active acoustic method, specifically comprises: 101) sending a sound signal to an underwater fish body and obtaining an echo signal scattered from the underwater fish body; 102) performing a filtering and an amplitude normalization preprocessing on the collected echo data of the fish body; 103) extracting features of the preprocessed echo signal with a specific method being as follows: segmenting the echo signal into a plurality of subsegments and extracting time-domain centroid of each subsegment as a feature quantity; and 104) inputting the feature quantities to a classifier to perform a classifying process and thus fish identification of the underwater fish body is completed. A self-adaption segmentation strategy or a uniform segmentation strategy can be adopted in the segmenting method of the step 103). Besides, the self-adaption segmentation strategy or the uniform segmentation strategy is adopted in the invention to segment the echo signal into the plurality of subsegments.

The invention discloses a remote sensing image classification algorithm based on mRMR selection and improved FCM clustering. The method comprises the following steps: S1, constructing a new object confidence coefficient (OC) index to measure an object-oriented multi-scale segmentation algorithm of a matching degree between any region and a geographic object, segmenting an image to generate image spots, and improving over-segmentation and under-segmentation problems in an image segmentation process; S2, introducing an mRMR feature selection algorithm, utilizing mutual information to measure thecorrelation and redundancy of different features, and searching a feature subset according to the information difference and the information entropy, so that the minimum redundancy exists between theselected feature and the target category, and the feature redundancy problem is solved; S3, performing feature distance calculation by adopting an improved FCM classification algorithm of significantfeature difference fusion to realize image classification; and S4, evaluating a pattern spot classification result of the remote sensing image. The images are classified, so that accurate classification of ground object categories is realized. According to the invention, the overall precision of classification reaches 94%.

The invention discloses an image feature matching method used for image retrieval, and belongs to the technical field of computers. By using the image feature matching method, an image is described from different types of four features including an image direction distribution feature, an image length distribution feature, image internal structure comparison feature distribution statistics and animage four-block gray feature; and finally image feature matching is performed by using a 7-dimensional eigenvector expression image feature. An algorithm solves the problems of large calculation amount, low efficiency and low accuracy of image matching in the prior art, greatly improves the image retrieval speed and performance, makes up for the limitation of a single feature, and can improve thematching accuracy.

The invention discloses a non-intrusive household appliance load identification method based on adaptive feature selection, which is specifically implemented according to the following steps: data preprocessing: denoising a selected REDD data set; performing event detection on the processed data through an improved generalized likelihood ratio test; extracting multi-dimensional load characteristics of the detected event change points; carrying out segmentation according to the power, and for multi-dimensional characteristics, using an adaptive mRmR algorithm to extract characteristics corresponding to each segment; taking the selected characteristics as load marks, and establishing a load feature library through an improved K-means algorithm; and on the basis of the load characteristic library, using the kNN algorithm to identify the load working state of the household electrical appliances in the user. The problem that in the prior art, misjudgment is likely to happen is solved, and the recognition accuracy is improved.

The embodiment of the invention provides an information processing method and device, electronic equipment and a storage medium, and can be applied to the field of artificial intelligence and the field of big data. The method comprises the steps: obtaining credit card application information of a to-be-evaluated user; processing the credit card application information by using the information evaluation model, and obtaining an evaluation result for the user, wherein the evaluation result is used for representing the possibility that the user has fraudulent behaviors, the information evaluationmodel is generated by training based on the credit card application information of the sample user of the first feature dimension, and the credit card application information of the sample user of the first feature dimension is obtained by performing dimension reduction processing on the credit card application information of the sample user of the second feature dimension.

The invention relates to the technical field of high beams, in particular to a deep learning-based high beam identification method, which comprises the following steps of: extracting characteristics of center brightness, edge symmetry, linear gradient and the like for representing light spot and halo image information, and judging the type of a vehicle lamp; extracting and classifying vehicle lampcategories based on a multi-target convolutional neural network, and achieving multi-task learning by using a deep learning algorithm; combining an improved DBSCAN algorithm and a least square methodto carry out fitting on the vehicle lamp category; predicting an inter-frame displacement limit value to adaptively generate a tracking search area, and optimizing a target matching function of adjacent frames to achieve multi-vehicle lamp tracking; after the high beam lamp is tracked for a certain distance, performing left and right lamp pairing through colinear constraint, and completing high beam lamp identification. Stray light interference is avoided, and the feature redundancy is reduced. Meanwhile, the single vehicle lamp is independently matched up and down, so that the influence of vehicle lamp refitting is avoided, and the recognition effect is remarkably improved by using a deep learning algorithm.