30results about How to "Improve classification recognition ability" patented technology

The invention belongs to the technical field of underwater acoustic signal processing, and particularly relates to an underwater acoustic signal target classification and recognition method based on deep learning. The method comprises the following steps: (1) carrying out feature extraction on an original underwater acoustic signal through a Gammatone filtering cepstrum coefficient (GFCC) algorithm; (2) extracting instantaneous energy and instantaneous frequency by utilizing an improved empirical mode decomposition (MEMD) algorithm, fusing the instantaneous energy and the instantaneous frequency with characteristic values extracted by a GFCC algorithm, and constructing a characteristic matrix; (3) establishing a Gaussian mixture model GMM, and keeping the individual characteristics of theunderwater acoustic signal target; And (4) finishing underwater target classification and recognition by using a deep neural network (DNN). According to the underwater acoustic signal target classification and recognition method, the problems that a traditional underwater acoustic signal target classification and recognition method is single in feature extraction and poor in noise resistance can be solved, the underwater acoustic signal target classification and recognition accuracy can be effectively improved, and certain adaptability is still achieved under the conditions of weak target acoustic signals, long distance and the like.

The invention discloses a neural network image classification and recognition method based on an optimized KPCA algorithm, and the method comprises the steps of calculating the cosine similarity of different vectors in a high-dimensional space, carrying out the dimension reduction of an original matrix of the KPCA algorithm through matrix rank minimization, reserving the effective information of original data to the maximum degree, and extracting better feature vectors to serve as weight values of convolutional layers, therefore, the problems that when an original KPCA algorithm is used for convolutional neural network image classification prediction, convolution kernel initialization calculation is complex, dimensionality disasters are likely to be caused, reliable features cannot be extracted, a whole network is difficult to train, and a network architecture is sensitive to image noise are solved. Therefore, the robustness and prediction performance of the whole network model are improved, and the effect of image classification and recognition is finally improved.

The invention discloses a grassland ecological data monitoring and intelligent decision-making integrated solution system, which comprises an ecological environment data acquisition module, a target area video data acquisition module, a sensor module, a data processing module and a cloud service platform module, the system can establish a cloud ecological database for a large-scale grassland ecological environment in a short time, and can perform a series of scientific analysis and decision-making on the ecological state of a target area from multiple dimensions in combination with big data and artificial intelligence application, including plant variety identification, ecological disaster early warning and plant disease and insect pest degree analysis. A user can take reasonable improvement measures while fully knowing the ecological state of the target area, valuable scientific research data is provided for scientific research personnel, and powerful technical support is formed for restoration and protection of the grassland ecological environment.

The invention discloses an audio scene recognition method combining deep neural network and a theme model as well as a system thereof. The method trains the audio event classification DNN neural network, a PLSA theme model, and an audio scene recognition DNN neural network at a training phase. In a testing phase, a test audio file passes through the audio event classification DNN neural network frame by frame; then an audio file-audio event co-occurrence matrix is constructed by the output of the neural network, the co-occurrence matrix is matrix-decomposed by the PLSA theme model, and decomposition is carried out to obtain the theme distribution of the test audio file on the potential theme; finally, the audio file-theme distribution is used as the input of the audio scene recognition DNNneural network, and the recognition result is obtained. The method innovatively combines the deep neural network with the theme model, and the introduction of the theme model is beneficial to providemore useful information for the deep neural network, and the method improves the classification and recognition capability of the network.

The invention discloses a rapid ship target detection method, a storage medium and computing equipment, and the method comprises the steps: establishing a feature pyramid full-convolution network anda double-branch module which sequentially comprise an input layer, a feature extraction layer, a feature fusion layer and an output layer, determining a ship data set, inputting a generated training set into the built feature pyramid full-convolution network, and generating a target detection result. According to the method, the ship in the image can be rapidly and accurately detected; and the detection result is accurate and rapid, the requirement on embedded equipment is low, and the method has very high practical application value in various aspects such as military affairs, civil use and the like.

The invention discloses an underwater acoustic communication signal modulation mode identification method based on an improved gating network and a residual network. The identification method comprises the following steps: receiving an underwater acoustic signal; improving the gated recurrent neural network to extract time features of the underwater acoustic signals; improving the residual neural network, and introducing a self-attention mechanism to carry out weighted reprocessing to obtain underwater acoustic signal spatial features; based on the correlation between the time features and the spatial features of the underwater acoustic signals, using an adaptive fusion strategy to fuse the obtained time features and the obtained spatial features of the underwater acoustic signals; sending the fused features to a full-connection neural network for network training; and after feature extraction and fusion are carried out on a to-be-recognized underwater acoustic signal, inputting the to-be-recognized underwater acoustic signal to the trained full-connection neural network for recognition, and finally outputting a recognition result. The technical scheme finally realizes underwater acoustic communication modulation mode intelligent identification with strong anti-interference and high accuracy.

The invention provides an electronic nose signal feature fusion method based on the separability degree and the dissimilarity degree and belongs to the technical field of electronic noise signal and information processing. The method comprises a step one of subjecting an electronic nose signal to feature extraction, a step two of performing feature selection and a step three of performing feature weighted fusion. The invention reserves classification information to the greatest degree while reducing dimensions and eliminating redundancy, and thus the classification identification rate is greatly improved, and the classification identification performance of an electronic nose is improved further.

The embodiment of the invention provides a method and a device for identifying safe dressing of an electric power worker. The method comprises the following steps: acquiring a dressing image of the electric power worker; and inputting the dressing image of the electric power worker into a preset dressing identification model to perform detection of the head and the body of the personnel, safety helmet classification identification and tool classification identification, and obtaining a safety helmet identification result and a tool identification result. According to the method and the device for identifying the safe dressing of the electric power worker provided by the embodiment of the invention, a lightweight network MobileNet is adopted as a basic feature extraction network, an optimized SSD target detection algorithm is combined, the head part and the body part of the personnel are preliminarily judged, and then different hierarchical features and different scale splicing features of the network are applied to different classification identification tasks. And the calculated amount is reduced, and meanwhile, the safety dressing classification and identification performance is improved.

The invention discloses an unbalanced big data set-oriented unsupervised text topic related gene extraction method, which comprises the following steps of: obtaining a clustering cluster of a high-dimensional sample set by adopting factor analysis and a density peak algorithm, and labeling an unlabeled sample; improving a CHI statistical matrix-based feature selection method by utilizing the average local density and the information entropy, and forceening the feature expression degree of low-density and small sample clusters; a fast fixed point algorithm based on negentropy is adopted, high-order statistical correlation between multi-dimensional data is analyzed, independent implicit topic feature genes are extracted, and removal of high-order redundancy between components is completed. Large-scale labeled samples do not need to be adopted for training, so that the pre-definition of a sample category relationship and a feature structure can be effectively avoided; and the influence ofan over-sampling or under-sampling method on the category distribution of the original unbalanced data set is overcome. The performance of the CHI statistical selection method is improved by correcting the feature category structure; effective feature dimension reduction under the condition of keeping the sample set recognition capability is also realized.

The invention discloses a remote sensing image target detection model establishment method based on context enhancement and an application, and belongs to the technical field of image processing, andthe method comprises the steps: building a to-be-trained target detection model based on a neural network, and carrying out the target detection and training on a remote sensing image, obtaining a remote sensing image target detection model based on context enhancement; in the target detection model, using each module for extracting a multi-scale feature map F of the remote sensing image; extracting global context information of the F to obtain M; respectively enhancing boundary information and category information in the F to obtain M<E> and M<E><cl> and respectively capturing informationassociation between channels in M<E> and M<E><cl> to obtain channel weights W<d> and W<c>; and fusing the M and M<E> according to the W<d> to obtain a boundary information enhanced feature map F<E>, fusing the M and M<E><cl> according to the W<c> to obtain a category information enhanced feature map F<E><cl>, and fusingF, F<E>, and F<E><cl> to obtain a feature map F<E><ct>, and carrying outtarget detection on the feature map F<E><ct>. The method can improve the target detection precision of a remote sensing image.

The invention discloses a target application program updating method and device based on feedback information. The method comprises the steps that actual scene information input in a target application program and result feedback information corresponding to the actual scene information are obtained; according to the actual scene information and the result feedback information, a target neural network model in the target application program is updated, the target neural network model is a neural network model obtained by conducting model training on a pre-trained neural network model through a training sample set, the training sample set comprises at least two sets of sample scene information, the different groups of sample scene information correspond to different actual sample recognition results of the target object, the loss function used during model training corresponds to different weights in the different groups of sample scene information, and the weights corresponding to the different groups of sample scene information are in negative correlation with the number of the sample scene information in the different groups of sample scene information.

The invention discloses a face recognition method and system. The method comprises the steps of: extracting feature vectors of face images with labels and face images without labels to obtain a firstfeature vector set and a second feature vector set; calculating soft label information of the label-free face image; calculating inter-class divergence and intra-class divergence of the face image training set and executing a linear discriminant analysis algorithm to obtain a discriminant projection matrix; performing dimension reduction on the feature vector of the to-be-detected face image, thefirst feature vector set and the second feature vector set by utilizing the identification projection matrix to obtain respective low-dimensional feature vectors, and inputting the low-dimensional feature vectors into a collaborative representation classifier to obtain respective collaborative representation codes, and calculating a reconstruction residual error by utilizing the collaborative representation code corresponding to each type, wherein the type of label with the minimum reconstruction residual error is a to-be-detected sample label. According to the method, the label information with the label data is transmitted to the data without the label, the number of the training samples with the label is increased, all the samples are identified and analyzed, and the CRC precision and discriminability are improved.

The invention discloses a bone marrow cell classification and identification method and system based on multiple features and multiple classifiers. The method comprises the following steps: training and testing bone marrow cells based on three transfer learning classifiers; respectively extracting a texture feature LBP, a shape feature HOG and a color feature HSV of the bone marrow cell image data set, performing fusion to obtain a feature fusion image, and performing fusion on the three migration classifiers by using a Keras model fusion algorithm to obtain a fusion classifier; training the fusion classifier by using the feature fusion image with the best test accuracy in the single classifier to obtain a multi-feature multi-classifier fusion model with the best test accuracy; positioning and segmenting bone marrow cells in the bone marrow cell image to obtain a model application data set; and testing the data set by using the multi-feature multi-classifier fusion model with the best test effect to obtain a final model application effect. The advantages of different features and different classifiers are combined, and the test classification accuracy of the bone marrow cells is improved.

The invention discloses a zero-sample image classification method based on an unknown class similar class set, and the method comprises the steps: carrying out the class prediction of an unknown classthrough a class classifier trained through the features of a known class sample, and obtaining the similarity degree of the unknown class and the known class on an image sample feature level; obtaining a similarity relation between the categories from an attribute level by utilizing a relation between the categories of the data set and the attributes; fusing the feature similar category set and the attribute similar category set of the unknown category to generate a similar category set of the unknown category; taking similar known class image samples of unknown classes as training samples toretrain class classifiers of the unknown classes, representing relationships between the classes and attributes by using the distinction degree of the attributes to the classes, and adding feature similarity relationship weights of the unknown classes and the known classes and relationship weights of the unknown classes and the attributes on the basis; and completing category prediction based onthe indirect attribute prediction model. According to the invention, the zero-sample image classification process is more objective and intuitive.

Aiming at the limitation of the prior art, the invention provides a litchi disease and insect pest identification method based on deep learning. According to the method, a deep learning training result is applied, various litchi diseases and insect pests can be automatically identified, and the problems of low efficiency, poor identification effect and the like of a traditional artificial disease and insect pest identification method are solved; according to a litchi disease and insect pest recognition model used in the scheme, an attention mechanism SimAM is introduced on the basis of a lightweight convolutional neural network ShuffleNetV2, an activation function Hardswitch is used, and a Dropout regularization method is adopted in a full connection layer; the litchi disease and insect pest recognition model can effectively extract important features, suppress interference of non-important features and improve network classification recognition performance, the number of network model parameters is not additionally increased, and storage and calculation expenses of the model are reduced.

The invention discloses a multi-source distillation-migration mechanical fault intelligent diagnosis method based on high-order moment matching, and the method comprises the steps: building a multi-source data set through the operation data collected from a plurality of mechanical devices, carrying out the preprocessing, and dividing the multi-source data set into a source domain data set, a target domain training data set, and a target domain test data set; constructing a multi-source distillation-transfer learning network model based on high-order moment matching, and performing high-order moment matching, maximum classifier difference and multi-source distillation training by using the source domain data set and the target domain training data set; and taking the target domain test data set as test input, and synthesizing outputs of the plurality of classifiers by using an adaptive weighting strategy to complete cross-domain fault diagnosis. According to the method, features of a source domain and a target domain are aligned at domain and category levels by utilizing multi-source data, the classification capability of the model on target samples is improved through multi-source distillation, and adaptive weighting is provided to integrate diagnosis results, so that the problem that the performance of a traditional method is reduced in cross-domain diagnosis is solved, and the performance of a deep model is greatly improved.

The invention discloses a reservoir computing hardware implementation method and device based on a coupled MEMS resonator, and the method comprises the steps: carrying out the preprocessing of a to-be-detected time sequence signal, so as to enable the to-be-detected time sequence signal to correspond to a virtual node of the coupled MEMS resonator; designing a nonlinear vibration equation of the coupled MEMS resonator, and regulating and controlling the coupled MEMS resonator to a preset nonlinear working point according to the equation; respectively detecting two signal test ends of the MEMScoupled resonator to obtain a first output signal and a second output signal corresponding to the to-be-tested signal corresponding to each moment, and feeding back the output signal corresponding tothe to-be-tested signal at the current moment to the virtual node corresponding to the next moment through a bidirectional time delay feedback loop; and performing regression training on the preset target value and the first output signal and the second output signal corresponding to the to-be-measured signal corresponding to each moment to obtain a weight coefficient required by calculation of the storage pool. According to the method, the data mapping dimension and the memory performance are enhanced, and richer nonlinear characteristics are provided for reserve pool calculation.

The invention relates to the technical field of image recognition, in particular to a distance-adaptive thermal infrared face recognition method, which comprises the following steps: constructing and training an improved thermal infrared image super-resolution enhancement network Retinex-CNN, and processing a thermal infrared image by using the trained improved thermal infrared image super-resolution enhancement network Retinex-CNN; on the basis of prior information obtained by near-infrared, different feature algorithms are utilized, local feature extraction is carried out on near-infrared and processed thermal infrared images, full feature fusion is carried out on the extracted different features, dimension reduction processing is carried out on the fused features, and then classification and recognition are carried out on the features after dimension reduction. Through the recognition method, the problem that face recognition is difficult due to the fact that thermal infrared imaging is greatly influenced by the distance when the distance changes can be effectively solved.

The invention provides a speaker identification method and system, and belongs to the technical field of speaker identification. The method comprises the steps of analyzing voice signals by using a confirmation network, extracting voiceprint features of the voice signals, and confirming whether the voice signals are from the same speaker or not; and analyzing the voiceprint features of the voice signals by using an identification network, and identifying the identity of a speaker of the voice signals. According to the method, specific voiceprint features of the speaker are extracted, channel noise interference is reduced, and the identification accuracy of the speaker is improved; the speaker confirmation network and the speaker identification network can assist each other by adopting a multi-task learning form, so that the confirmation effect and the identification effect are improved; a human brain thinking mode is simulated by combining a speaker confirmation technology, and identification is carried out by combining the confirmation technology under the condition that the number of speakers is large and the identification difficulty is high; and through the simulation application of the human brain thinking model, the intelligence of the algorithm is improved, and the identification ability of the model under the condition of high difficulty is enhanced.

The invention relates to a method and a device for adjusting an equivalent Q value of a silicon microresonator RC (Resistance-Capacitance) system by utilizing mechanical pumping. The method comprises the following steps of: determining a resonator modal frequency and a mechanical pumping signal; the equivalent Q value is dynamically adjusted through mechanical pumping to verify the experiment; further performing complex nonlinear response on the pre-processed to-be-tested signal by using a coupling MEMS resonator RC system under specified conditions according to a verification experiment; filtering and collecting a response signal of the resonator; accurate control, data detection, downsampling, logical operation and the like of a digital delay feedback control program; and training and testing the test data set. According to the method, energy exchange between different modes of the resonator is achieved through mechanical pumping, and dynamic adjustment of the equivalent Q value of the used modes is achieved under the condition that it is guaranteed that the system works under the strong non-linear rich condition (the high vacuum environment). According to the method, the nonlinear mapping performance of reservoir calculation is ensured, and the calculation speed is increased.