31results about How to "Reduce redundant features" patented technology

The invention discloses a DDNN based on cloud-edge collaborative computing and its construction method and application, which are applied to object classification under multi-view images of the Internet of Things. DDNN includes: an edge side, which adopts a feature bag model to extract multiple features The similarity of the image is measured and the histogram vector of each view is obtained statistically; the feature weighted fusion is performed on the histogram vector of each view, and the classification of the edge side target is obtained based on the fused histogram vector. If the classification accuracy is not enough, the extracted The feature images of each viewing angle are transmitted to the cloud; the cloud is used to carry out feature weighted fusion of all feature images, perform convolution and classification operations on the fused feature images, and obtain the target classification in the cloud. The present invention introduces a feature bag model on the edge side to reduce the amount of parameters; in addition, multi-view weighted feature fusion reduces multi-view redundant features and increases the effectiveness of feature expression capabilities. The invention reduces the amount of parameters of DDNN and the traffic of cloud-edge communication and improves the overall performance of DDNN.

The invention discloses a local sparse representation object tracking method based on LO regularization. Different from traditional L1 object tracking methods, the method herein proposes the combination between a LO norm and a structural local sparse appearance model, fully utilizes sparse coding, better differentiates objects from backgrounds in the course of tracking, and models local shielding and the like interferences by using a trivial template, and further increases the robustness of noise interference in the course of tracking. In order to enable an object model to better deal with continuous changes of the appearance of an object in the course of tracking, according to the invention, the method, through the construction of an object model set, adopts the LO norm in reconstructing the object in the course of tracking and adopts probabilistic policy in replacing a certain template in the object template set with a reconstructing result, thus realizing the update of template dynamics and further increasing the stability of the algorithm. According to the invention, the method which is directed at the NP problem of the object function optimization solution based on the LO norm, employs the APG algorithm in realizing effective solution.

The invention discloses a cross-dataset facial expression recognition model construction method and a recognition method. The reconstruction error of the training expression image is constrained by an adaptive non-negative weighting matrix, which strengthens the role of important features in image data representation and reduces the need for significant reconstruction. Useless features for errors. In addition, projecting the training set into an appropriate subspace through the mapping matrix can better reveal the intrinsic similarity between image samples across datasets, so that subspaces based on low-rank and sparse representations can learn robust image reconstructions. In order to facilitate the final cross-dataset facial expression recognition.

The present invention proposes a new optimal feature subset selection method based on the complementarity of classification capability structure vectors, aiming at the classification capability evaluation criteria in which single values ​​are used as features or feature subsets in most existing methods. This method defines the feature classification ability structure vector and the complementary features of the classification ability structure vector in binary form, and uses the dichotomy method to calculate the threshold of feature classification and discrimination ability in each subcategory problem, and on this basis, according to the different characteristics of the selected feature subset The principle of structural complementarity maximization and greedy strategy are used to select the optimal feature subset. This method not only fully considers the different evaluations of each feature on the classification ability of different categories, but also follows the principle of maximizing the structural complementarity of the classification ability in the process of feature selection. It not only conforms to the natural law of complementary advantages, but also maximizes the feature classification information, so as to obtain a better feature subset, effectively reduce redundant features, and improve the accuracy of classification prediction.

The present invention provides a feature selection method and device. The method includes: extracting M features from the obtained image; using a filtering model algorithm to filter the M features to obtain N features; using the first The encapsulated model algorithm searches the N features to obtain a first feature subset, and uses the second encapsulated model algorithm to search the N features to obtain a second feature subset; for the first feature Subset and the second feature subset are recombined to obtain a third feature subset including K features, and the best feature of the image is selected from the first feature subset, the second feature subset and the third feature subset Excellent feature set. Through the technical solution of the present invention, the most effective features can be selected from a large number of features, the optimal feature subset can be obtained, the computational complexity can be reduced, the classification accuracy and classification performance of face recognition can be improved, the recognition ability of faces can be improved, and the recognition ability of faces can be improved. recognition rate and improve information security.

The invention discloses a social network text sentiment analysis method based on a deformable self-attention mechanism, which is used for analyzing the sentiment of user speech. The steps are as follows: segment each sentence in the user utterance text data into words, and represent each word with a word vector; input the word vector sequence into a bidirectional recurrent neural network (Bi-LSTM) to obtain the coded representation of each word; Using the deformable self-attention mechanism, the coded representation of words is converted into a variety of sentence coded representations with different context ranges; multiple sentence coded representations are fused to obtain a sentence coded representation; the fused sentence coded representation is input to the feedforward neural network Classify in the network (FFN), output the result; minimize the cross-entropy loss function iteratively train the model to update the parameters according to the model output result and the real data result; input the social network text to be classified into the trained model to obtain the sentiment analysis result.

The invention discloses an intrusion detection method based on semi-supervised learning. The implementation steps include: selecting a mixed sample set initially containing labeled samples and unlabeled samples to be tested, calculating the information gain of each feature value in the feature space and completing the process based on The feature selection of information entropy, and then screen the labeled samples based on the feature selection of information entropy, and use the new training data obtained by screening for the semi-supervised training of the classifier based on LapSVM, and use the trained classifier to be tested. The samples are classified, the best evaluation value of the detection index is determined according to the detection index, and the classification result corresponding to the evaluation value of the best detection index is output. The present invention adopts the method of feature selection to deal with the redundancy phenomenon that is easy to appear in the network environment data, and uses a small number of labeled samples and a large amount of unlabeled data to create a semi-supervised learning model, while reducing the false alarm rate and improving the detection rate, It can reduce data redundancy and improve detection efficiency.

The invention relates to a plane feature fusion method in the process of SLAM mapping, belonging to the field of image processing. The steps are: 1) According to the point cloud data of the vehicle lidar, segment the plane features in the lidar coordinate system at different times; 2) Obtain the vehicle GPS position and IMU heading angle information at different continuous times, so as to obtain the translation matrix relative to the initial position and rotation matrix; 3) Transform all plane feature information in the lidar coordinate system to the world coordinate system; 4) Angle size for the intersection line between any plane feature and the x‑o‑y plane at any two different times The calculation of the calculation of the coincidence degree of the feature circle and the calculation of the relative distance; 5) Delete the class with a small number of features in the class, and find the corresponding linear regression equation for the features of the remaining classes, and then construct the world coordinate system. global map. The invention reduces redundant features, improves the simplicity of map expression and the precision of map construction.