64results about How to "Improve classification recognition accuracy" patented technology

The invention discloses a method for classifying and recognizing capsule network images of an improved reconstructed network. The method includes: S1, constructing a capsule network; 2, inputting an image training set to that capsule network, wherein the capsule network complete image classification recognition calibration after training and learning; 3, inputting an image to be classified to thatcapsule network, wherein the maximum value of the output vector vj of the work network is the obtain recognition result; S4, the capsule network outputting the recognition result of the image to be classified, wherein the reconstructed network structure of the capsule network is deconvolution operation. The beneficial effects are: a new reconstructed network structure is provided, in which the vector is restored to the image by deconvolution operation, the network parameters are adjusted by comparing the error between the restored image and the original image, the computational parameters arereduced, and more running memory is freed for the hardware equipment.

The invention discloses an object-oriented remote sensing image coastline extraction method which belongs to the field of full automatic remote sensing image coastline extraction. The steps are as follows: carrying out filtering treatment on a remote sensing image; selecting the segmentation algorithm to segment the remote sensing image; utilizing sample points for classifying the remote sensing image after segmentation; utilizing the seed growth method for extracting a seawater region; and finally utilizing the corresponding judgment criterion to extract a coastline. By utilizing the method disclosed by the invention, the accuracy of the existing coastline extraction algorithm can be improved, the correct rate of seawater extraction after segmentation and classification treatment throughthe object-oriented method is improved and the accuracy of the coastline extraction is ensured. As the coastline changes frequently, the method can provide convenience for maintenance and updating ofcoastline information of all levels of national fundamental geographic information databases.

The invention discloses a wheat field weed detection method based on deep learning, and the method comprises the steps: collecting a large number of wheat and wheat field main weed pictures at different growth stages, building a data set, and dividing the data set into a training set and a test set; inputting the training set into a preset convolutional neural network model for training through atransfer learning method to obtain a crop weed classification recognizer, and testing the crop weed classification recognizer by using a test set to obtain a classification recognition result so as toperform fine adjustment; generating a large number of interest domains with different sizes on the to-be-detected picture by adopting a sliding window method, and inputting each interest domain intoa crop weed classification recognizer for classification and recognition to obtain a corresponding prediction category and a correct probability; and screening out an interest domain corresponding tothe local maximum correct probability of each type from all interest domains by applying a non-maximum suppression algorithm, and outputting a classification and positioning prediction result. According to the method, crops and weeds can be quickly and accurately identified and positioned, and the requirement for data is low.

The invention discloses a width transfer learning network and a rolling bearing fault diagnosis method based on the same, and belongs to the technical field of bearing fault diagnosis. The invention provides a novel width transfer learning network and a rolling bearing intelligent diagnosis method based on the same, and aims to solve the problems of scarcity of vibration data with mark informationof a rolling bearing under a variable load, large distribution difference between source domain data and target domain data in the same state, unbalanced distribution of multi-state data and low diagnosis accuracy and model training efficiency. According to the invention, a width learning system is utilized to extract features of source domain data and target domain data and construct a sample set, and on the basis, a balanced distribution adaptation method in transfer learning is adopted to reduce the difference between a source domain and a target domain. A chicken swarm algorithm is introduced to optimize width transfer learning network parameters and establishing a width transfer learning network model. The proposed network model is applied to rolling bearing fault intelligent diagnosis under the variable load, and an experimental result verifies the high efficiency and accuracy of the proposed method.

The invention belongs to the technical field of big data intelligent analysis, and particularly relates to an incremental learning method and system based on a small number of labeled samples. Expanding and enhancing a small number of labeled samples to obtain a reliable label data set, and learning the network by using the reliable label data set to obtain a pre-training model; based on a networkpre-training model, performing prediction classification on a large number of unlabeled samples, and constructing an incremental learning candidate data set; combining the reliable label data set andthe incremental learning candidate data set to obtain an incremental learning data set, performing incremental learning on the network pre-training model, and performing calibration learning on the incremental learning model by using the reliable label data set; and carrying out prediction classification on the unlabeled data by utilizing the pre-training model after calibration learning, and judging return and re-execution by setting loop iteration conditions. According to the method, reliable sample data used for classification and recognition are obtained through incremental learning underthe condition that only a small number of labeled samples exist, and the classification and recognition performance and accuracy are improved.

The invention discloses an aerial handwriting recognition system and a method based on machine vision. The handwriting character video input part is used for collecting the aerial handwriting character video of the input specific color in real time and generating the trajectory picture from the collected trajectory points. The pre-processing part is used to filter, binarize and pre-process the trajectory image generated by the trajectory points. The character segmentation part is used to convert the RGB color space of the track image generated by the pre-processed track points into HSV color space, and then the characters in HSV are segmented out by a GA-Otsu segmentation algorithm; the feature extracting part extracts the features of the segmented characters. The classification and recognition part is used to get the optimal SVM classification model by PSO and SVM, and output the recognition results of all characters. The invention only needs to capture the writing track of a specificcolor, has high efficiency, saves cost, is convenient to carry, and does not need to judge the start, end and end of writing.

The invention discloses a spectrum matching method based on spectrum curve waveform similarity. The method includes the following steps: 1) manufacturing samples to be measured; 2) collecting spectrums of all test samples; 3) calculating first derivatives of original spectrums; 4) eliminating a zero value in each first derivative and utilizing a non-zero first derivative of an adjacent zero-value first derivative to replace the zero-value first derivative value; 5) conducting calculation and statistics to obtain the average value of specific values of all first derivatives of the same band of two spectrums and utilizing the average value as the matching degree of the two spectrums. By means of the method, the curve waveform similarity is adopted as the spectrum matching index to judge the similarity between the spectrums, dependence of the spectrum matching method on absolute strength of the spectrums is reduced. By means of the method, higher classification recognition accuracy can be achieved, and the method has great significance on improvement of spectrum database searching speed and spectrum database searching accuracy and is favorable for spectrum data information sharing and research result popularization.

The invention discloses a fatigue driving early warning system based on electroencephalogram analysis. The fatigue driving early warning system includes an electroencephalogram acquisition subsystem,an electroencephalogram analysis subsystem and a driving early warning subsystem, wherein the electroencephalogram acquisition subsystem used for acquiring electroencephalogram analog signals, obtaining and transmitting electroencephalogram digital signals by preprocessing; the electroencephalogram analysis subsystem used for extracting feature and analyzing of the electroencephalogram digital signals and conducting time domain waveform display, the electroencephalogram digital signals are recognized by using an SVM multiple-classification recognition algorithm, and the classification result of driving state dispersed grade which a driver is located is obtained and transmitted; the driving early warning subsystem includes an intelligent early warning hand ring used for conducting early warning control according to the driving state dispersed grade which the driver is located and a vehicle-mounted monitoring device used for conducting early warning control according to the driving statedispersed grade in which the driver is located. The fatigue driving early warning system uses the electroencephalogram signals to classify and recognize, whether the driver monitored by the vehicle-mounted monitoring device is in a heavy dispersed state or not judged, early warning is conducted, and the fatigue driving early warning system is high in judging precision, timely in early warning, convenient to carry and easy to operate and has the high practical value.

The invention relates to the field of weed identification, in particular to a broad-spectrum crop weed identification and positioning method based on improved deep learning. The method comprises the following steps: collecting pictures of a small amount of target crops and field weeds in different growth stages; expanding a small amount of picture data into a training set and a test set on a largescale by adopting a sample enhancement technology; extracting a main area of the image; proposing a target crop feature extraction method based on a pre-training network InceptionV3, and carrying outsecondary training on the initialized model by using a new data set by adopting an improved transfer learning method on the basis of the trained model to obtain a target crop and weed recognition model; proposing an activation function for improving the network; and finely adjusting the model parameters by utilizing the test set. By adopting the method, the target crops and weeds can be quickly and accurately identified and positioned, the requirement on the acquired image data is relatively low, and the demand on the data is flexible.

The invention discloses an active voice detection method based on noise scene recognition, and the method comprises the steps: extracting an optimal feature facing a noise classification task from anaudio signal, inputting a feature value into a noise type classifier, so as to recognize a noise type in the audio signal, according to the noise type, determining a preferred feature and a classifiersuitable for the voice and noise classification task, extracting optimal features oriented to the voice and noise classification task from the audio signals, inputting the optimal features into a voice and noise classifier, and judging whether the voice signals exist in the audio signals or not. According to activity voice detection system based on noise scene recognition, before binary classification of noisy voice and noise signals is carried out, according to the method, the current noise type is detected and identified, the most distinctive feature combination is preferably selected for the specific noise type, model parameters can be designed for the specific noise type, and the effectiveness and stability of the performance of the whole detection process under different noise typesare ensured.

The invention belongs to the technical field of text classification, and particularly relates to a text classification method and system fusing a self-attention mechanism and deep learning, and the method comprises the steps: obtaining a to-be-classified text data set, and carrying out the preprocessing; and carrying out classification processing on the preprocessed to-be-classified text data set by utilizing a trained deep learning model, wherein the deep learning model comprises an ERNIE pre-training module used for extracting sentence-level word vector representation in a to-be-classified text data set, a BiLSTM module used for extracting context information of each word according to the sentence-level word vector representation, a DPCNN module which is used for deeply extracting the context information of each word according to the sentence-level word vector representation and the context information of each word, an attention mechanism module which is used for extracting a text depth information distribution weight according to the sentence-level side vector representation and the context information of each word, and a Softmax module which is used for carrying out classified output according to the text depth information distribution weight. The text classification recognition effect is improved by combining an attention mechanism and each model.

The invention relates to a magnetic shoe surface defect detection system and method, and relates to the field of magnetic shoe surface defect detection. The system comprises an image acquisition device and a computer system. The image acquisition device is used for acquiring magnetic shoe surface images. The computer system performs suspicious area judgment on the magnetic shoe surface image and divides the suspicious area to acquire a suspicious area image. The computer system extracts geometric features, gray features and texture features of the suspicious area image. The computer system inputs the geometric features, the gray features and the texture features into a classifier which is trained in advance. The detection result of the magnetic shoe surface defects is obtained, high-precision classification and identification of the magnetic shoe defects are realized through a classification and identification algorithm, automatic detection of the magnetic shoe defects is realized, thelabor cost is reduced, and the efficiency and the identification accuracy are improved.

The invention discloses a feature processing method and device for point cloud data based on rotation and terminal equipment. The feature processing method comprises the steps: obtaining target point cloud data; extracting a rotation invariant feature from the target point cloud data by using a rotation mapping module; carrying out multi-dimensional feature processing on the rotation invariant features by utilizing a plurality of clustering modules, wherein the clustering modules are sequentially connected according to a clustering number sequence from large to small; and classifying the rotation invariant features processed by the multi-dimensional features by using a classifier module to obtain a classification result of the rotation invariant features. According to the invention, the rotation invariant features extracted from the target point cloud data are processed by combining the rotation mapping module and the plurality of clustering modules; the robustness of target recognition based on the 3D point cloud data can be ensured; the classification recognition precision of the target point cloud data is enhanced; and meanwhile the number requirement for training data of the deep learning model and the calculation cost during training of the deep learning model are reduced.

An EEG signal feature dimension reduction method based on weighted principal component analysis comprises: extracting samples of m EEG signals of fatigue driving and dividing the samples into a training set and a test set for training, and obtaining the total classification accuracy A and n different classification accuracies; subtracting the total accuracy rate A from different classification accuracy rates to obtain n difference values; normalizing the n difference values to obtain n weights; constructing a weight diagonal matrix for the n weights; writing samples of the m EEG signals into an m * n-dimensional matrix; multiplying the m * n-dimensional matrix by the diagonal matrix of the weight to obtain weighted EEG signal feature data; calculating and decomposing a covariance matrix toobtain a characteristic value of the covariance matrix and a unitized characteristic vector corresponding to the characteristic value; selecting unitized feature vectors corresponding to the featurevalues of the first k covariance matrixes to be combined to form a mapping matrix; and thus, obtaining dimension-reduced EEG signal characteristic data. According to the invention, the classificationidentification precision is effectively improved and the training time of the identification model is reduced.

The invention provides a fine-grained cervical cell image three-stage identification method, which comprises the following steps: firstly, denoising a whole cervical cell image acquired by electronicimaging equipment to improve the image quality, removing a background to determine a foreground cell region, and segmenting out cell nucleuses and cytoplasm; secondly, extracting shape, color, textureand abstract features from the segmented cell regions, and carrying out selective fusion on the features by adopting a multi-kernel learning method so as to optimize a feature mode; and finally, removing non-cervical cell impurities according to the optimized characteristic mode, and performing multi-classification on the cervical cell images by adopting a cost-sensitive active learning method. The identification process system is complete, can extract medical features of multiple types of cervical cell images, performs selective fusion optimization on feature modes, removes impurities by using a cost-sensitive learning method, solves the problem of unbalanced dichotomy, and can effectively improve the accuracy of cervical cell classification and identification.

The invention discloses a self-adaptive neural network training method for image recognition. The method comprises the steps: acquiring and preprocessing an image data set; constructing a convolutional neural network model, and setting an adaptive loss function; inputting the preprocessed image data into a convolutional neural network model for forward propagation to obtain a feature vector and a classification layer weight of the image; calculating an adaptive loss function according to the image feature vector and the classification layer weight, and judging whether the convolutional neural network model converges; performing back propagation on the convolutional neural network model according to the adaptive loss function, and updating the weight of the classification layer; and progressively increasing the number of iterations, and updating the adaptive loss function. Compared with a softmax-based classification loss function training method, the method provided by the invention has the advantages that the setting of hyper-parameters is reduced, the convergence speed of the convolutional neural network model can be accelerated, and the image classification and recognition accuracy can be improved.

The invention discloses a face parallel identification method based on deep learning and Spark, and via the method, face identification is realized via a deep learning framework and Spark parallelization. In the image feature extraction process, a graph is distributed to different cluster nodes via Partition, each node of a Spark cluster is used to train parameters and models in parallel, one nodeis generated in each cluster node, the models are used for parallel face identification later, and less time is consumed via parallelization. In the face identification process, Spark characteristicsare used, the characteristic is broadcast to each cluster node in a parallel identification manner, calculation is carried out in each node, and each node generates a corresponding result of the faceto be identified. The results of all sub nodes are gathered in a main node, and are compared to obtain a final result.

The invention is applicable to the technical field of computers, and provides a sample screening and expression recognition method, a neural network, equipment and a storage medium. The method comprises the following steps: utilizing a multi-component sample; training neural networks, scooter for each iteration, determining a first sample distance between the anchor sample and the positive sampleand a second sample distance between the anchor sample and the negative sample, according to the distribution statistical characteristics of the sample distances; constructing a boundary condition forscreening the multi-tuple sample; the boundary condition being used for screening the multi-element group samples, and the reserved result obtained by screening entering the training of the next iterative step, so that the abnormal multi-element group samples can be screened out in the training process of the neural network, the influence of the abnormal multi-element group samples on the training result of the neural network is avoided, and the expression classification and recognition accuracy is improved.

The invention belongs to the technical field of circuit breakers, and discloses a control system and method of an intelligent permanent-magnet vacuum circuit breaker. The control system comprises a signal acquisition module, a closing control module, a central control module, a photoelectric isolation module, a fault identification module, an alarm and a display screen. The closing time of the circuit breaker can be compensated by the closing control module, so that the opening and closing time of the circuit breaker can be kept stable under different working conditions, and a reliable basis is provided for realizing synchronous closing. And meanwhile, the fault identification module combines a support vector machine with incremental learning to identify common faults of the permanent-magnet mechanism high-voltage vacuum circuit breaker. Small sample data is fully used for training a model, the generalization ability of fault recognition is improved, a sample library is continuously enriched by learning new data, and the classification recognition accuracy can be improved to more than 90%.

The invention relates to the field of graph classification and recognition, in particular to an image classification and recognition method and device based on a self-adaptive dynamic convolutional network and computer equipment, and the method comprises the steps: obtaining a to-be-detected image, inputting the to-be-detected image into a preprocessing block, and obtaining a shallow feature graph and graph parameter information of the image; combining the image parameter information obtained after preprocessing with the to-be-detected image, and inputting the combined image parameter information and the to-be-detected image into an adaptive dynamic convolutional network of a backbone network to obtain image global features; wherein the adaptive dynamic convolution is to select a convolution kernel with a corresponding shape according to the corresponding parameter information; inputting an image shallow layer feature map obtained after preprocessing into a branch network, and extracting local features of the to-be-detected image; and carrying out feature fusion on the local features and the global features, inputting the fused features into a classification network, and outputting classification identification information of the to-be-detected image. According to the method, the calculation cost required for image classification and recognition is low, the precision is high, and the applicability of related products is high.

The invention discloses an idiodynamic artificial limb system based on brain-computer hybrid intelligence. The system comprises a visual stimulation module, a portable electroencephalogram signal acquisition device, an electroencephalogram signal processing and recognition module and an artificial limb which are connected in sequence; the visual stimulation module is used for providing visual stimulation options for a testee; the electroencephalogram signal processing and recognition module is used for collecting SSMVEP electroencephalogram signals of a testee; and the electroencephalogram signal processing and recognizing module is used for conducting feature extraction and classification on the SSMVEP electroencephalogram signals and driving the artificial limb to conduct six actions according to the classification result, wherein the six actions comprise pressing elevator buttons, lifting a thumb, pouring water into a cup feeding food into the mouth, moving an object upwards to a cabinet, and picking up an object by two fingers. According to the method, classification and identification precision of a cerebral apoplexy patient in a fatigue state can be improved, the misoperationrate of the cerebral apoplexy patient on the artificial limb in a fatigue state is further reduced, the daily life of the cerebral apoplexy patient is facilitated, and applicability of the idiodynamic artificial limb system is improved.

The invention discloses a ground crawler-type unmanned vehicle control method based on an arm myoelectric signal, which overcomes the problems of lack of flexibility, equipment complexity and vulnerable to external environmental interferences existing in the prior art. The method comprises the following steps: step 1), correctly wearing an electromyography sensor and performing initialization setting: (1), enabling a user to wear the electromyography sensor in the middle position of the hand and the elbow, so that an electrode is tightly attached to the skin, thus a reference electrode can beattached to the ulnar flexor of wrist, thereby ensuring that the myoelectric signal can be collected; (2), performing an initialization operation, and firstly, initializing the electromyography sensorand a crawler-type unmanned vehicle, thereby testing whether the reliable wireless communication connection is established or not through a signal processing device and the crawler-type unmanned vehicle in a certain area; step 2), collecting the myoelectric signal and identifying different gestures and transmitting the gestures to the crawler-type unmanned vehicle; step 3), parsing a command by the crawler-type unmanned vehicle and executing an action of the command; and step 4), wirelessly transmitting contents shot by a video camera, transmitting the contents back to the signal processing device, and displaying the contents in a computer.

The invention discloses a high-resolution image and machine learning-based high-altitude region crop classification and identification method, which comprises the following steps of: screening out an optimal feature combination through a recursive feature elimination strategy based on a random forest by using a domestic GF6-PMS satellite image and combining features such as spectrum, texture, vegetation index and topographic factor; and calculating a Gini index to obtain an importance score of each input feature, and further utilizing a two-layer stack-driven integrated classification model (including three single classifier models of Random Forest, XGBoost and AdaBoost) to classify and identify the crops in the high-altitude region. The Stacking model constructed on the basis of the optimal feature combination (Green, Red, NIR, TVI, GNDVI, BlueMean, GreenMean, RedMean, NIRMean, DEM) can improve the classification and recognition precision of crops in high-altitude areas to a large extent, especially the classification and recognition precision of bulk crops with large planting areas, and provides a scientific reference basis for crop remote sensing recognition of domestic high-resolution satellite images in high-altitude areas.

The invention provides an electroencephalogram motion intention recognition method and system. The method comprises the steps: obtaining an electroencephalogram data set based on a motion intention; after the electroencephalogram data set is subjected to feature extraction of at least two different dimensions, splicing and classifying the obtained features of all the dimensions, obtaining a motion intention recognition model with the highest classification accuracy, and outputting a classification result; the accuracy of motion intention classification and recognition can be effectively improved.

The embodiment of the invention provides a classification model training method. The method comprises steps that the gradient contribution corresponding to each sample in the current batch of data is obtained in a training process of a classification model; samples with the gradient contribution larger than or equal to a preset gradient contribution threshold value in the current batch of data serve as first difficult samples and are added into a difficult sample set, the difficult sample set comprises second difficult samples, and the second difficult samples are samples with the gradient contribution larger than or equal to the preset gradient contribution threshold value in the non-current batch of data; and a third difficult sample is selected from the difficult sample set according to a preset screening rule, and the classification model is trained according to the third difficult sample. According to the method, difficult sample mining is carried out on the current batch of data and the non-current batch of data, and the third difficult sample is screened out from the first difficult sample and the second difficult sample, so the screening range of the third difficult sample is enlarged, more representative difficult samples can be obtained to train the classification model, and classification recognition accuracy of the classification model is improved.

The invention discloses a forestry pest recognition and detection method based on pooling vision Transform. The method aims at discriminating and detecting different types of forestry pests in different periods. The method mainly comprises the steps: reasonably dividing data sets of different periods and different types into a training set and a verification set; expanding training set pictures by using a sample enhancement technology; proposing a forestry pest recognition feature extraction method based on a pre-trained PiT neural network model, adopting an improved transfer learning method to perform secondary training on an initialized model by using a new data set on the basis of the trained model to obtain a forestry pest recognition classifier; and performing fine adjustment on model parameters by using the test set. By adopting the method, the types and periods of the forestry pests can be quickly and accurately recognized, and the requirement on the acquired image data is low.

The invention discloses an image classification and identification method based on the regional bicubic interpolation technology, and belongs to the field of image processing. An existing image processing method has the problem that the image classification identification rate is reduced due to image distortion caused by scale transformation. According to the image classification and identification method based on the regional bicubic interpolation technology, the method comprises the steps of extracting the features and K-of a training sample through SIFT; performing means clustering statistics on the word frequency, and training a classifier model after training data is obtained in the classification step; subjecting the images of a test sample to identical zooming-in and zooming-out changes through regional bicubic interpolation, and then carrying out retrieval and classification through the BOF algorithm. According to the invention, the classification and identification accuracy ofdistorted images caused by zooming can be improved.

The invention discloses a remote sensing image lake ice classification and recognition method based on a neural network. The method includes: firstly obtaining a remote sensing image of an annual winter ice period from a database, and performing radiometric calibration processing on the obtained remote sensing image to obtain a preliminary remote sensing image; performing FLAASH atmospheric correction on the obtained remote sensing image after radiometric calibration processing, and obtaining the final processed remote sensing image; constructing a BP neural network model; training the neuralnetwork model constructed in the step (3), and testing the trained neural network model to determine whether the accuracy requirement is satisfied, going to the step (5) if so, and returning to reconstruct the neural network model if not; and adding the final processed remote sensing image to the constructed neural network model, and classifying lake ice in the remote sensing image. The classification accuracy obtained by the invention is high.

The invention provides a GNN-based ship classification and identification method and device, and the method comprises the steps: extracting the features of ship AIS data, and constructing a total sample set which is a three-dimensional matrix; converting the total sample set into graph structure data, and dividing the total sample set into a training set and a test set; and training a GNN network model through the training set, inputting features of ship AIS data of all samples to be tested in the test set into the trained GNN network model to test effectiveness of the GNN network, and classifying the ships to be classified through the GNN network passing the test, wherein the GNN network model is a GNN neural network model with two layers of graph convolution layers. According to the scheme of the invention, spatial features can be effectively extracted for machine learning by using the ship trajectory, and the classification and recognition accuracy of the ship trajectory can be improved.