8results about How to "Good classification accuracy" patented technology

The invention discloses an SAR ship target classification method based on deep dense connection and metric learning, and mainly solves the problems of inaccurate feature extraction and poor classification effect in the prior art. According to the scheme, the method comprises the following steps: 1) acquiring ship target SAR image training data, and expanding the ship target SAR image training data; 2) establishing a network model composed of a deep dense connection layer and an embedded conversion layer; 3) sending the expanded training data into the network constructed in the step 2), and using cross entropy loss with L2 norm regular terms to perform preliminary training on the network; 4) adding the triple loss and a Fisher discrimination criterion-based regular term into the loss function in the step 3), and sending training data to continue to train the network model to obtain a finally trained network model; and 5) sending the test data to the trained network model to obtain a ship classification result. According to the method, deep feature extraction can be better completed, the classification performance is improved, and the method can be used for sea area ship monitoring and target classification.

The invention discloses a domain adaptive privacy protection method based on differential privacy for a deep neural network. A novel deep network framework is provided, and data privacy can be protected while a domain adaptation technology is realized. In the real scene of transfer learning, such as schools and hospitals, a training data set is generally private, and a scheme for flexibly protecting domain adaptation technical privacy does not exist nowadays, so that the method has very strong practicability. According to the method, domain adaptation training is carried out by using the ideaof adversarial learning, and privacy protection is carried out on the domain adaptation training process through differential privacy for the first time. Experimental results show that the model can complete domain adaptation tasks with ideal accuracy under proper privacy consumption.

The invention relates to a deep neural network space spectrum classification method for a high-spectral image and belongs to the technical field of deep learning and high-spectral remote sensing image classification. In the method, grouped space spectral features are used as input, according to input grouping features, a regularization item is added to an optimization target at a first layer of a deep neural network, and extraction of the space spectral features and waveband selection are realized. The method takes algorithm features of a deep belief network into consideration, also takes features of space information into consideration, performs individual processing on space groups of each waveband and is different from a deep convolutional network in which parameters in a convolutional nucleus are the same; and the algorithm can automatically attenuate weights of wavebands having quite small classification effects, realizes adaptive feature extraction and waveband selection, can obtain better classification accuracy compared to the typical deep belief network and has wide application prospect.

The invention discloses a feature selection method based on a binary quantum particle swarm algorithm. The feature selection method carries out feature correlation analysis by using a maximum information coefficient, carries out feature selection processing through an improved BQPSO algorithm, and carries out classification accuracy verification by using an SVM. Experimental results of gene expression profiles show that feature selection based on the improved BQPSO algorithm is a feasible method. According to the feature selection method, a standard binary quantum particle swarm optimization algorithm is mainly improved, and a mode based on a complete learning strategy is used for calculation of a local attractor, and meanwhile the diversity of particle swarms is improved by introducing the variation thought of a genetic algorithm. Experiments show that the improved BQPSO algorithm is used for feature selection, and better classification accuracy can be obtained.

The invention discloses a gas sensor fault mode recognition method based on multi-scale analysis and ensemble learning, and the method comprises: carrying out the multi-scale analysis of a gas sensorfault signal, obtaining time sequences under different scale factors, and respectively calculating the weighted permutation entropy of each time sequence to form a composite multi-scale weighted permutation entropy feature vector; performing dimensionality reduction on the composite multi-scale weighted permutation entropy through a Fisher discrimination method, and enabling the composite multi-scale weighted permutation entropy to serve as a fault feature sample of pattern recognition; and constructing a plurality of base learners by using an ensemble learning method, carrying out classification prediction on the sub-sample sets of the fault feature sample set, and then summarizing classification results of the plurality of base learners to obtain a gas sensor fault mode identification result. According to the method, the difference of different fault types can be highlighted, the selected ensemble learning classifier has more excellent generalization performance and better classification accuracy for gas sensor fault recognition, and serious accidents are avoided.

The invention relates to an emotion discrimination method based on fine-grained annotation data. The method comprises the following steps: collecting financial news data, dividing news data into a labeled sample set and an unlabeled sample set; training a first classifier and a second classifier through the labeled sample set and the unlabeled sample set; the first classifier can screen out the key sentences in the article; the second classifier judges the emotional tendency of the article; model parameters of the first classifier and model parameters of the second classifier are obtained respectively; and adding the data with high confidence in the classification result into the annotation sample set, selecting the most worthy annotated data C from the unannotated sample set by using an active learning theory, sending the most worthy annotated data C to a worker for annotation, and circularly training the emotion discrimination model until the classification precision is achieved, andending the training to obtain the discrimination model.

The invention provides an improved multidimensional scaling heterogeneous cost-sensitive decision tree building method, which comprises the steps of selecting splitSi from a candidate attribute according to a target function f(Si) of an attribute Si, extending branches meeting the condition splitS=splitSi from a node, supposing that the number of the branches meeting the condition is k and adding a blank node to the node, namely determining the number of the branches of the current node to be k+1; and simultaneously carrying out pruning operation on leaf nodes by using a first pruning technology, carrying out pruning while building the tree, and stopping building the tree when two conditions are met as follows: (1) Yi is supposed to be a sample set meeting the condition splitS=splitSi in a training dataset, if Yi is null, one leaf node is added and the sample set is marked as the most common type in the training dataset; and (2) all examples in the node belong to the same type. According to the method provided by the invention, the classification accuracy is improved; the bias problem in the classification process is solved; and multiple cost impact factors and the blank node in the branches of a decision tree are considered, and the next step of classification operation can be continued through the blank node if an unknown classification result does not conform to a current model.

The invention belongs to the field of radar data processing, and provides an air flight target classification and identification method based on radar track data. The method comprises the following steps: carrying out data preprocessing on flight path data of an air flight target collected by a radar, and eliminating the influence of data differences caused by different dimensions of data on a training model; and then reducing the data dimension of the obtained flight data through principal component analysis, extracting key data features, and improving the modeling speed and precision. The time sequence of the flight data is considered, a model is established for the data through the recurrent neural network, the test data is classified and identified, and the good classification performance is obtained. According to the invention, additional hardware is not needed, and classification and identification of radar aerial flying targets such as light and small unmanned aerial vehicles, birds, helicopters and civil aircraft are realized in a single radar antenna period.