44results about How to "Effective feature extraction" patented technology

The invention discloses an objective positioning and classifying algorithm based on deep learning. The algorithm comprises the following steps: S1, inputting a picture into a first network and outputting a series of objective positioning frames and scores; S2, inputting a picture and a series of sub windows into a second network; S3, forwardly propagating the networks to a last convolution layer to generate a characteristic pattern; S4, performing coordinate transformation on the sub windows by using a zooming coefficient in order to map the coordinates to the characteristic pattern; S5, acquiring a characteristic from the characteristic pattern by using the zoomed sub windows and pooling the characteristic to a regular size; S6, classifying the pooled data to obtain a classification result and a score of an area; and S7, inputting the objective positioning frames and the classification result of the area into a classifier to carry out classification, and outputting the type and the coordinates of the objective. The algorithm may extract characteristics, classifies the objectives, performs positioning and identification by using a unified network, and greatly increases calculating speed.

The invention relates to a character-level nested deep network-based text classification method. The method comprises the following steps of S1, constructing a character vector matrix table; S2, performing short text preprocessing; S3, extracting high-dimensional sequence features by improved Resnet; and S4, performing LSTM network classification. The character level-based text conversion can effectively perform conversion on all texts; compared with a conventional vector space model, the dimension is reduced remarkably; all the texts can be effectively converted; and low-frequency words are not ignored. In addition, the improved Resnet can self-learn a feature extraction method; and compared with conventional methods such as a TF-IDF formula, a mutual information quantity, an informationgain, an x2 statistical quantity and the like, the extracted features are more effective and abstract. Finally, the LSTM network classification can consider a sequence relationship between the words,so that the text classification can be performed more accurately.

The invention discloses a multi-modal fusion saliency detection method based on a convolution attention module. The method comprises: in a training stage, constructing a convolutional neural network;inputting the left viewpoint image and the depth image of the original image into a convolutional neural network for training to obtain a corresponding saliency detection image; calculating a loss function between a set formed by saliency detection images generated by the model and a set formed by corresponding real human eye gaze images to obtain an optimal weight vector and an offset term of theconvolutional neural network classification training model; and in a test stage, inputting the three-dimensional image in the selected data set into the trained convolutional neural network model toobtain a saliency detection image. According to the visual saliency detection method, extraction of image features is optimized by applying a novel module, multi-scale and multi-mode feature fusion iscarried out, and finally the detection efficiency and the detection accuracy of visual saliency detection are improved.

The invention discloses an upper ocean thermal structure prediction method based on a deep belief network. The prediction method comprises the following steps of 1) determining proper environment parameters as input factors and predication values of a thermal structure prediction model; 2) establishing an upper ocean thermal structure sample data set, dividing sample data into training data and test data, and performing unified data preprocessing; 3) establishing the deep belief network, and performing non-monitoring pre-training on the sample data layer by layer to obtain relatively excellent parameters of the model preliminarily; 4) based on a back propagation algorithm, performing fine tuning on the parameters of the model according to marks of the training sample to determine optimal parameters; and 5) applying an established upper ocean thermal structure prediction model to test data, and obtaining a predicated temperature value at a specific depth of the upper ocean from the output layer. By adoption of the upper ocean thermal structure prediction method based on the deep belief network provided by the invention, the problems of single fitting and over-fitting of the conventional method are overcome; and a characteristic relation between the ocean surface layer environment parameters and the upper ocean thermal structure is effectively extracted, so that the upper ocean thermal structure prediction accuracy is improved.

The invention provides a training method and device for a brain tumor recognition model based on semi-supervised learning. The brain tumor recognition model comprises a detection network and a classification network, and the method comprises the following steps of obtaining a first training sample set and a second training sample set; performing unsupervised learning on the detection network and the classification network by using the first training sample set to generate a pre-trained detection network and a pre-trained classification network; training the pre-trained detection network and the pre-trained classification network by using the second training sample set to generate a trained detection network and a trained classification network; and outputting a trained brain tumor recognition model, wherein the trained brain tumor recognition model comprises the trained detection network and the trained classification network. According to the method, fine marking and coarse marking data are fully utilized in a semi-supervised learning mode, so a more robust deep convolutional neural network is obtained.

The invention relates to a hyperspectral image classification method, and concretely relates to a hyperspectral image classification method based on spectral space attention fusion and a deformable convolutional residual network. The objective of the invention is to solve the problem of low classification accuracy of hyperspectral images caused by insufficient spectrum and spatial feature extraction and overfitting under small samples due to the fact that the hyperspectral images contain abundant information in the existing hyperspectral image classification. The method comprises the following steps: 1, acquiring a hyperspectral image data set and a corresponding label vector data set; 2, establishing an SSAF-DCR network based on spectrum space attention fusion and a deformable convolution residual error; 3, inputting the x1, the x2, the Y1 and the Y2 into a network SSAF-DCR, and performing iterative optimization by adopting an Adam algorithm to obtain an optimal network; and 4, inputting x3 into the optimal network to carry out classification result prediction. The method is applied to the field of hyperspectral image classification.

The invention belongs to the technical field of aero-engines, and particularly relates to an aero-engine service life prediction method based on improved LSTM. Comprising the steps of 1, processing original sensor data acquired by a sensor, and constructing a training sample which comprises a training set and a test set; step 2, on the basis of constructing the training sample in the step 1, constructing an LSTM structure model as an engine residual life prediction model; and 3, inputting the test set in the step 1 into the LSTM structure model constructed in the step 2 to obtain a predicted RUL value, and evaluating the obtained predicted RUL value by adopting RMSE and Score evaluation indexes. According to the aero-engine life prediction method based on SDAE and LSTM, the advantage of unsupervised feature extraction of a deep encoder is utilized, effective feature extraction is performed on an engine sensor signal, low efficiency of manual feature extraction and prediction uncertainty caused by the low efficiency are avoided, and the prediction accuracy is improved. And the residual life of the engine is predicted by using the advantage of processing the time sequence data by the LSTM model.

The invention discloses an image super-resolution implementation algorithm based on an information filtering network. The algorithm comprises a feature extraction module, a plurality of superposed information filtering modules and a reconstruction module, wherein the feature extraction module extracts features from a low-resolution image firstly, then superposes a plurality of information filtering modules, extracts residual information step by step, and finally, the reconstruction module aggregates the obtained high-resolution residual feature representation to generate a residual image; andelement-level item-by-item addition operation is performed on the residual image and the up-sampled low-resolution image. According to the image super-resolution implementation algorithm based on theinformation filtering network, due to the fact that the number of filters of each layer is relatively small, and group convolution is used, the network algorithm has the advantage of being high in execution speed.

A human body posture recognition method using a millimeter-wave radar comprises the steps that echo data of the millimeter-wave radar in the detection range of the millimeter-wave radar are acquired, and the echo data of the millimeter-wave radar comprise point cloud information; determining that a human body exists in the detection range according to the millimeter wave radar echo data; determining the mass center of the human body according to the point cloud information of the millimeter wave radar echo data; noise of the point cloud information is removed according to the mass center of the human body; and recognizing a human body posture according to the denoised point cloud information. The invention further provides a computer readable storage medium, privacy leakage is avoided, and the accuracy of human body posture recognition is improved.

The invention discloses a no-reference image quality evaluation method based on an adversarial generative network. The method comprises the steps of firstly preprocessing an image to obtain a similargraph-SSIM _ MAP corresponding to a distortion graph, then training a neural network framework based on a densenet network, inputting the distortion graph into the trained network to obtain the similar graph of the distortion graph, and obtaining a corresponding quality score through the similar graph. Innovations are made for a generator and a loss function in a network. The method comprises thefollowing steps: firstly, in a generation network part, adopting a 60-layer densenet network framework; in the discrimination network part, adopting a simple classification network; and in the loss function part, adopting a mode of adding L1 norm loss to the cross entropy of a discriminator; and finally, iteratively training a relatively good generation network model, and generating a similar picture for an output picture, namely a distorted picture, through the network.

The invention discloses a no-reference image quality map generation method based on an adversarial generative network. According to the method, a U-net network framework with eight down-samples and eight up-samples is adopted in a generation network part; a classification network is adopted in the discrimination network part; a mode of adding L1 norm loss to the cross entropy of a discriminator isadopted in the loss function part; and finally, a generative network model is iteratively trained, a similar graph of the input distortion graph is acquired through the generative network model, anda corresponding quality score is obtained through the similar graph. The method has no reference quality evaluation. Quality evaluation is performed on the distorted image under the condition of no natural image by using the trained neural network framework. The problem of calculating the quality score of the similar graph with the weight problem is solved. Based on an adversarial generative network and U-net, graph-to-graph conversion and migration are realized more effectively. An experiment result has a good result in graph-to-graph implementation, and the simulated mass fraction and the real mass fraction have strong correlation and small errors.

The invention relates to a relation extraction method for an automatic knowledge graph construction system ; the method comprises the steps: coding and converting a text into word vectors, and preliminarily extracting text features; generating a syntactic dependency tree by using a syntactic dependency structure of the text, weighting each relation category to generate a weighted dependency adjacency matrix, and extracting syntactic dependency information in the text by using a graph convolutional neural network; synchronously, directly acting a multi-head attention mechanism on the encoded text to generate an attention matrix, and extracting information except syntactic dependency information of the text by using a graph convolutional neural network with the same structure; and finally, obtaining feature representations of the two entities and sentences, scoring all possible relation categories by using a feedforward neural network and a normalized exponential function, and selecting a relation with the highest score as a relation classification result. According to the method, information of different dimensions of the text can be fully acquired, and an excellent effect is achieved on a public data set for relation extraction.