141results about How to "Reduce the risk of overfitting" patented technology

The invention relates to an electric power telecommunication network device fault prediction method based on improved LSTM. The invention provides a data preprocessing and time sequence input construction method for the first time. Compared with the simple recurrent neural network, the LSTM is more likely to learn long-term dependence and can solve the prediction problems related to sequences. Because of the strong association between device alerts, the independence of variables can be ensured through PCA. The target replication strategy is also used for improving the LSTM, which can bring local error information in every step. Compared with a simple target output only in the last step, the strategy can improve the accuracy of the model and reduce the risk of over fitting. In combination with dropout, the invention proposes a prediction model of LSTM, which can achieve better prediction precision by deep learning. At the same time, the LSTM is usd for modeling electric power telecommunication network alarm data for the first time and identifying the timing sequence mode therein.

The invention discloses a human brain function network classification method based on a convolution neural network, and belongs to the field of brain science research. the method comprises the following steps: obtaining resting state fMRI data and preprocessing the data; generating simulation data; performing data set partitioning; performing classification of human brain functional network basedon convolution neural network. The method of the invention is based on a convolution neural network, and uses Element-wise Filters to give unique weights to each edge and node of the human brain functional network data, thereby constructing a multi-layer neural network comprising an edge-to-node layer and a node-to-graph layer. The method of the invention can better utilize the topological structure information of the data of the human brain function network and carry out the feature expression, thereby improving the classification effect, and the method is reasonable and reliable, and can provide powerful help for the diagnosis of neuropsychiatric diseases.

The invention discloses an extreme learning machine-based hyperspectral remote sensing image ground object classification method. An original extreme learning machine network is expanded into a hierarchical multi-channel fusion network. In terms of network training, the method is different from the least squares algorithm-based output weight solving strategy of the original ELM (extreme learning machine) and the global iterative optimization strategy of a deep learning network; according to the method of the invention, a greedy layer-by-layer training mode is adopted to train a hierarchical network layer by layer, and therefore, the training time of the network is greatly shortened; and in the layer-by-layer training process, a l1 regular optimization item is added into the training solving model of each layer of the network separately, so that parameter solving results are sparser, and the risk of over-fitting can be lowered. In terms of network functions, A single-hidden layer ELM network focus on solving the fitting and classification problems of simple data, while the different levels of the network model provided by the invention achieve target data feature learning or feature fusion, the network model of the invention integrates the advantages of high training speed and strong generalization capacity of the single-hidden layer ELM network, and therefore, the in-orbit realization of the model is facilitated, and the requirements of emergency response tasks can be satisfied.

The embodiment of the invention discloses a video classification method, and is used for overcoming the defect in the prior art that videos cannot be accurately classified, and improving the degree of accuracy of video classification. The video classification method includes obtaining information in a video, the information in the video including image information, light stream information and acoustic information; utilizing a deep neural network to generate first reference information corresponding to the image information, second reference information corresponding to the light stream information and third reference information corresponding to the acoustic information; processing the video according to the first reference information, the second reference information and the third reference information to obtain a confidence degree matrix of the video and a category relation matrix of the video; and substituting the confidence degree matrix of the video and the category relation matrix of the video into an objective function to obtain an objective fusion parameter of the video, wherein the objective fusion parameter is used for classifying the video.

The invention discloses a video smoke identification detection method based on transfer learning. The method comprises the steps of generating analog data to enlarge the number of smoke image samples;preprocessing the image data set; constructing a target detection network; using an ImageNet image dataset to pre-training an improved VGG-16 network; training a target detection network on the labeled smoke data set by adopting a transfer learning mode, wherein the feature extraction network part uses improved VGG-16 network pre-training weights to perform feature initialization; and extractinga key frame from the video, inputting the key frame into the model for identification and detection, and if smoke is found, returning coordinate information and positioning an area thereof in the video image. Model performance under the condition of limited smoke data is improved by using a transfer learning technology, and a smoke area in a video image can be automatically identified and positioned.

The invention discloses a lightweight fine-grained image recognition method for cross-layer feature interaction in a weak supervision scene, and the method comprises the steps: constructing a novel residual module through employing multi-layer aggregation grouping convolution to replace conventional convolution, and enabling the novel residual module to be directly embedded into a deep residual network frame, thereby achieving the lightweight of a basic network; then, performing modeling on the interaction between the features by calculating efficient low-rank approximate polynomial kernel pooling, compressing the feature description vector dimension, reducing the storage occupation and calculation cost of a classification full-connection layer, meanwhile, the pooling scheme enables the linear classifier to have the discrimination capability equivalent to that of a high-order polynomial kernel classifier, and the recognition precision is remarkably improved; and finally, using a cross-layer feature interaction network framework to combine the feature diversity, the feature learning and expression ability is enhanced, and the overfitting risk is reduced. The comprehensive performance of the lightweight fine-grained image recognition method based on cross-layer feature interaction in the weak supervision scene in the three aspects of recognition accuracy, calculation complexity and technical feasibility is at the current leading level.

The embodiment of the invention discloses an unsupervised pedestrian re-identification method based on fuzzy depth clustering. The method comprises the following steps: extracting pedestrian image features by using pedestrian image feature extraction network model; extracting pedestrian image features by using the pedestrian image feature extraction network model; extracting pedestrian image features by using fuzzy depth clustering. Construct fuzzy depth clustering network and initialize it; The fuzzy depth clustering network is used to learn the new feature space and clustering center, and the fuzzy labels are assigned to the unlabeled pedestrian images. Using reliability samples to train the pedestrian image feature extraction network model; Alternate training until the reliability sample is saturated; Using the trained pedestrian image feature extraction network model to extract the test pedestrian image features, the unsupervised pedestrian re-recognition results are obtained by calculating the feature distance. The invention utilizes fuzzy depth clustering network to learn new feature space, which is favorable for clustering of complex pedestrian images and distribution of fuzzy labels, and utilizes reliability samples with fuzzy labels to train the feature extraction network, thereby reducing the risk of over-fitting and improving the correct rate of unsupervised pedestrian recognition and matching.

The invention belongs to the field of image target recognition and provides a remote sensing ship identification method based on dense feature fusion and pixel-level attention, and aims to solve the problems that a classical neural network easily identifies a plurality of dense targets as one target under a remote sensing image ship target identification task, a large number of small targets are missed, boundary frames are easy to overlap and the like. According to the main scheme, data set division is carried out on a remote sensing image data set to obtain a training set and a test set, anddata enhancement of the training set is carried out. RGB three-channel average values r < mean >, g < mean > and b < mean > of the original remote sensing image data set are calculated and the RGB three-channel values of the images are correspondingly subtracted in the expanded data set from the r < mean >, g < mean > and b < mean >; the obtained data set is input into an improved Faster RCNN network to be trained, core modules of the network are a dense feature fusion network and a pixel-level attention network, and the network outputs candidate rotation boxes and category scores of the candidate rotation boxes; and skew IOU-based rotating frame non-maximum suppression is carried out on the obtained result to obtain an identification result of the ship target in the remote sensing image.

The invention discloses a distributed optical fiber vibration signal feature extraction and identification method, which belongs to the field of optical fiber sensing signal processing, and comprisesthe following steps of: firstly, acquiring a space-time matrix signal of a vibration source, extracting a space column signal, dividing a short-time signal unit, and constructing an optical cable vibration event data set; constructing, training and optimizing an improved mCNN model, and performing feature evaluation on features extracted by the model during optimization until model iteration is optimal; secondly, extracting time structure feature vectors under multiple scales in parallel by utilizing an optimal mCNN model, recombining the time structure feature vectors into a short-time feature sequence according to a time sequence, and constructing a time structure feature sequence set; finally, constructing and training an HMM model, and constructing an offline vibration event HMM modellibrary to serve as a classifier for vibration source recognition. The problems that in the prior art, local structure features and time sequence features of distributed optical fiber vibration signals cannot be extracted at the same time, and the vibration source recognition accuracy and the generalization ability of the model are low are solved.

The invention discloses a hemodialysis treatment scheme auxiliary decision-making method based on an Attention-LSTM model. The method comprises the following steps: 1, preprocessing hemodialysis historical medical record data labeled by a professional dialysis doctor to obtain a total sample set; 2, establishing a long-short-term memory network based on a self-attention mechanism, and allocating attention weight alphat to each time sequence medical record of an input model so as to calculate a medical record information global feature vector c; 3, inputting the global feature vector c into a multi-task sharing layer to learn a scheme label; and 4, allocating a weight lambdan to the loss of each learning task, and taking the weighted sum of the loss as the total loss; and updating network parameters of the model by using an Adam optimization algorithm to obtain a final hemodialysis treatment scheme auxiliary decision-making model. The model can assist a dialysis clinician in making an individualized hemodialysis treatment scheme for a patient, the manual workload of the clinician is reduced, and the hemodialysis efficiency of a hemodialysis center is improved.

The application discloses a model training method and apparatus. The method comprises: acquiring a deep learning model of a source domain; acquiring training data of a target domain; adjusting a weight parameter of the deep learning model of the source domain by using the training data of the target domain to obtain a deep learning model of the target domain; with the deep learning model of the target domain, extracting data features of the training data; and training a K nearest neighbor classification model by using the data features to obtain an identification model. Therefore, the model training cost is lowered; the model accuracy is improved; and the over-fitting risk is reduced.

The invention provides a bullet screen text classification method, a bullet screen text classification device, equipment and a storage medium. The method comprises the steps: obtaining an imbalance training data set with a pre-marked category, and dividing the training data set into a sufficient sample and an insufficient sample; training the sufficient samples by adopting a textCNN model; carrying out model training on the insufficient samples by adopting an SVM classifier; inputting a text to be tested into the trained textCNN model, and outputting classification probabilities of various categories in sufficient samples; and if the output classification probability is smaller than a first preset threshold, inputting the to-be-tested text into a trained SVM classifier, and outputting a predicted category. According to the method, the classification models for different text scales are obtained through separate training according to the sizes of the training samples, then the two classification models are combined to be used for classifying the to-be-detected text, the problem of data imbalance of the training samples is solved, compared with single model training, the risk of over-fitting can be reduced, bias is reduced, and the recognition accuracy is higher.

The invention relates to a picture stylization processing method based on deep learning. The method includes steps: S1: establishing and training a neural network; S2: segmenting a to-be-processed picture into a plurality of ultra-pixels; S3: analyzing each ultra-pixel by employing the trained neural network, and marking the environment category with the highest adaption degree therefor; S4: defining the environment category with the maximum marking frequency of the ultra-pixel of all the ultra-pixels as the environment category of the picture; S5: extracting an object in the picture, and determining a target major object according to the position of the object in the picture; and S6: sharpening or blurring a background according to a stylization processing requirement, wherein the stylization processing requirement comprises background strengthening and background weakening. Compared with the prior art, the method is advantaged by fast processing speed and good effect.

The invention provides a pathological classification method and system based on multi-modal deep learning. The method comprises: extracting pre-selected attributes from the electronic medical recordsto serve as feature representation vectors of structural data, randomly discarding the feature representation vectors according to a preset proportion after being averagely amplified, and replacing the discarded parts with numbers 0 to serve as medical record feature vectors of the structural data in the electronic medical records; obtaining a histopathology image corresponding to the electronic medical record, performing global average pooling on the feature map of each convolutional layer of the convolutional neural network, and splicing the feature maps into a one-dimensional vector to serve as a rich image feature vector of the histopathology image; and splicing the image feature vector and the medical record feature vector together to obtain a multi-mode fusion vector, and inputting the multi-mode fusion vector into a full connection layer to obtain a binarized pathological classification result. The technical problem that the accuracy of pathological benign and malignant classification through single-mode feature representation is not high is solved.

The invention relates to an image number character recognition method. An algorithm structure is optimized, and influences of illumination and inclination are processed in an image preprocessing stage, thereby ensuring that a better binary character image is acquired before character recognition; then, a two-dimensional PCA algorithm is adopted to extract primary characteristics in an image so as to reduce the dimension and operation complexity of the image; a decision tree algorithm and an Adaboost algorithm are combined to optimize a classification algorithm, thereby ensuring the recognition efficiency and reducing the over-fitting risk; and finally, a nine-one division method is used for verifying the generalization ability of a classification model so as to ensure the stability of a classifier. By virtue of algorithm optimization, the accuracy rate and speed of number character recognition are obviously increased.

The invention provides a coronary artery stenosis detection method and device, computer equipment and a storage medium. The coronary artery stenosis detection method and device are characterized in that a detection result is obtained by inputting an obtained to-be-detected image into a coronary artery stenosis detection model; the coronary artery stenosis detection model comprises a trunk network,a segmentation network and a stenosis analysis network, and the output of the trunk network is connected with the input of the segmentation network and the input of the stenosis analysis network; thedetection result comprises a coronary artery segmentation result and a coronary artery stenosis result. Due to the structure of the coronary artery stenosis detection model, the coronary artery stenosis detection model comprises a segmentation network and a stenosis analysis network. Therefore, the coronary artery stenosis detection model is a multi-task model, can segment the input images to bedetected at the same time and conduct stenosis detection on the input image to be detected, so as to obtain detection results including at least two types, namely the coronary artery segmentation result and the coronary artery stenosis result.

The invention discloses a partial discharge signal pattern recognition method, device and equipment and a storage medium. The method comprises the steps of obtaining partial discharge data, carrying out the preprocessing of the partial discharge data, generating a data set of an adversarial network through employing the preprocessed data as a training condition, and dividing the data set into a training data set and a test data set; generating an adversarial network model and a machine learning classifier according to training target construction conditions; training a conditional generative adversarial network model in the training data set; inputting a random noise signal into a generator of the conditional generative adversarial network model to enable the generator to generate new partial discharge data, and obtaining a new data set according to the new partial discharge data; and after training the machine learning classifier in the new data set, verifying accuracy of the machinelearning classifier in the test data set. The method is advantaged in that the data set is expanded by generating the new data, so technical problems of difficult data sampling and low pattern recognition precision in the prior art are solved.

The invention discloses a pedestrian re-identification method based on self-excitation discriminative feature learning, which comprises the following steps: (1) selecting a pedestrian re-identification network, and adding a negative branch on the original network; (2) in the training stage, generating a classification loss function by the original network, generating a confrontation loss functionand a mutual exclusion response item between the original network and the negative branch to form an objective function, and utilizing a random gradient descent method to optimize the whole network; (3) in the test stage, removing a negative branch, only keeping the part, in front of the classifier, of the original network as a trained network model, and inputting a pedestrian picture for extracting a feature vector test; and (4) in the pedestrian retrieval stage, extracting the feature vector of each picture in the picture library by using the trained network model, and selecting the identityof the picture with the highest similarity with the feature vector of the to-be-queried pedestrian picture as a final recognition result. According to the invention, the effect of the existing pedestrian re-identification network can be improved.

The invention relates to a leukocyte image detection and recognition model construction method based on transfer learning and application. The leukocyte image detection and recognition model is trained by constructing a leukocyte image detection and recognition model based on transfer learning, taking a coco data set and labeled leukocyte images as samples. The leukocyte image detection and recognition model extracts three-level fused feature pyramids, different prior contour frames are set corresponding to leukocytes with different size features, network convergence is accelerated, and the recognition efficiency and accuracy are improved. The leukocyte images are automatically and efficiently detected, recognized and counted based on the leukocyte image detection and recognition model, the detection and recognition precision is improved, human and subjective factors are eliminated, and fairness and objectivity are achieved. According to the method, the reasoning time for detecting andidentifying one image is 0.05 second, and analysis of 20 leukocyte images can be completed per second. And the identification accuracy of six normal types of leukocytes reaches 98%.

The invention discloses an epilepsy detection method based on adaptive weighted feature fusion of a deep network stack model. The method comprises the following steps of: 1, after an original electroencephalogram signal is preprocessed, conducting discrete Fourier transform on data of each channel of the electroencephalogram signal, and obtaining an average amplitude spectrum of the electroencephalogram signal; randomly dividing the obtained average amplitude spectrum features into several parts so as to be suitable for a deep network stack model of adaptive weighted feature fusion; 2, performing second feature extraction on the data set obtained in the step 1 by using convolutional neural networks of different structures; and 3, performing adaptive weighted feature fusion on the features extracted by different convolutional neural networks, and finally predicting the category of the sample by an error correction output coding model based on a support vector machine. According to the invention, through a stack integration method and an adaptive weighted feature fusion algorithm, the system can fuse features extracted by depth networks of different structures, and the epilepsy prediction effect is improved.

The invention provides a video hash retrieval method of an attention mechanism. The method comprises the following steps: (1) video preprocessing: sampling video frames, and constructing video pairs;(2) video frame feature extraction: performing feature extraction on each frame by using a convolutional neural network; (3) video feature learning: learning the video by using a twin network, a long-term and short-term memory (LSTM) neural network and an attention mechanism; (4) dimension reduction and training: performing dimension reduction on the video features by using a full connection layerto obtain a hash code with a desired length, and learning network parameters by using a gradient descent algorithm; and (5) retrieval: obtaining the hash code of each video by using one path of network of the twin network, calculating the Hamming distance with other videos, and performing sorting to obtain the nearest video. Compared with the prior art, the method has the advantages that the spatial information and the time information of the video are learned at the same time, the calculation cost is greatly reduced through sampling and other technologies, and the accuracy of video retrievalis also improved.

The invention discloses a sample expansion method and system based on foreground and background feature fusion. The method comprises the steps: dividing a remote sensing ground feature classificationdata set into a source data set and a target data set based on ground feature categories; constructing a small sample source ground object classification task based on the source data set, and training a feature extractor, a hybrid model and a classifier based on the small sample source ground object classification task; constructing a small sample target ground object classification task based onthe target data set; performing sample expansion by using the trained feature extractor and the hybrid model based on the target classification task; wherein each task comprises a first task and a second task; the mixed feature is a feature synthesized by a foreground feature and a background feature by using a mixed model; according to the method, the hybrid model is trained based on the classification task, additional manual annotation is not added to expand the training sample, the training cost is reduced, the trained feature extractor and the hybrid model are utilized to expand the target data set, the classifier is trained, and the sample expansion method is realized.

The invention discloses a fault diagnosis method based on an improved convolutional neural network, and relates to the technical field of fault diagnosis. The method includes: establishing an improvedconvolutional neural network model, wherein the improved convolutional neural network model sequentially comprises an input layer, a plurality of feature extraction layers, a newly added convolutional layer, a full connection layer and an output layer, each feature extraction layer sequentially comprises a convolution layer and a pooling layer; performing training by using the training set and the test set based on an improved convolutional neural network model to obtain a fault diagnosis model for automatic fault diagnosis. According to the method, a newly-added convolution layer is furtherarranged between a feature extraction layer and a full connection layer, the newly-added convolution layer can extract deep features of the model, the generalization ability of the model is effectively enhanced, meanwhile, the feature extraction layer is improved, the risk of over-fitting can be reduced to a certain extent, and the calculation speed is increased.

The invention discloses a photovoltaic station generated power prediction method and system. The method comprises the steps: building a training sample set according to the historical power data of aphotovoltaic station and the meteorological data of a corresponding time period, and distributing a sample weight to each training sample in the training sample set; training a random forest model byadopting a sub-training sample set generated from the training sample set through a self-service sampling method, and calculating an error rate and a weight coefficient of the random forest model under the current sample weight according to an adaptive enhancement algorithm; under a preset iteration frequency, updating the sample weight according to the error rate and the weight coefficient, sequentially training random forest models, and weighting the random forest models according to the weight coefficient to obtain a weighted random forest prediction model; and predicting the meteorologicaldata in the to-be-predicted time period by adopting the weighted random forest prediction model to obtain the power generation power of the photovoltaic station. Information in multi-dimensional features is fully mined, the uncertainty problem existing in photovoltaic power generation is solved, and the reliability and accuracy of photovoltaic station power prediction are improved.

The invention discloses a stereoscopic image visual salience extraction method based on a dual-learning network, which forms a training set of a human gaze map, a left viewpoint color image and a leftparallax image of the stereoscopic image. Then, based on the training set, the depth learning model is constructed by using the feature extraction technology of VGG network model. Secondly, the depthlearning model is trained by using the human gaze map as the monitor and the left viewpoint color image and left parallax image as the input parameters. Then, the left view color image and the left parallax image of the stereoscopic image to be significantly extracted from the vision are taken as input parameters and input into the model obtained from the training to obtain the visually significant image of the stereoscopic image to be significantly extracted from the vision; the advantage is that it can run quickly, and has strong robustness and prediction accuracy.

The embodiment of the invention discloses a medical image classification device and system. The device comprises a classifier pre-construction module and an image classification module. The classifierpre-construction module is used for extracting a plurality of ROI images of each medical sample image in the training sample set, and then each class of ROI images is trained by utilizing a three-dimensional convolutional neural network to obtain corresponding base classifiers; and an optimal base classifier is selected from the plurality of base classifiers by using a genetic algorithm based ona preset fitness value, and finally an image classifier is generated based on each optimal classifier. The image classification module is used for inputting the to-be-identified medical image into a pre-constructed image classifier to obtain a classification result of the to-be-identified medical image. According to the device and sytem, the accuracy of medical image classification is improved, the medical image of the early-stage Alzheimer's disease can be accurately recognized, and assistance is provided for prevention and diagnosis of patients with the early-stage Alzheimer's disease.

The invention discloses a sow delivery time prediction system based on late pregnancy posture conversion characteristics, which is characterized by comprising an image acquisition module, a network transmission module, a video storage unit, a local server, a central computing module and a mobile terminal receiving module, the image acquisition module acquires a sow video in a post-pregnancy deadline column environment, and the acquired data is stored in the video storage unit; and the screenshot image frames of the local server are uploaded to the central computing module through the network transmission module. According to the invention, end-to-end automatic operation and control are realized from the image acquisition module to the final mobile terminal receiving module, the sow management efficiency can be greatly improved, and the death risk of piglets is reduced.

The invention discloses a device, a method and a system for monitoring a state of a hydraulic system. The method comprises the following steps that supervised learning model training is provided, andspecifically, a training set and a test set are divided according to an existing sample data set, and training, verification and optimization processes of the supervised learning model is completed; and the system is initialized, and specifically, the system is started. According to the device, the method and the system for monitoring the state of the hydraulic system, the method comprises the following steps that firstly, data dimensionality reduction is conducted on the large-scale hydraulic measurement original data by using an unsupervised PCA algorithm, the data processing amount is greatly reduced, the training and prediction speed are remarkably improved, the overfitting risk is reduced, the training and prediction speed are remarkably improved, the generalization ability of a modelis improved, and good prediction accuracy is achieved; and the requirement for real-time accurate evaluation of the hydraulic system is met, finally, the prediction accuracy of the hydraulic state can be remarkably improved, and a better application prospect is brought.

Provided is a traffic flow prediction method based on a corrosion de-noising deep belief network. Traffic flow prediction refers to the prediction of the current and future traffic flow according to the traffic flow at historical moments. A random corrosion layer is proposed as a means of regularization to relieve the interdependence of some neurons, improve the generalization ability of the prediction model and reduce the risk of over-fitting. Meanwhile, a traffic flow prediction model is established based on specific application scenarios and the spatial correlation and temporal regularity of traffic flow, to realize accurate, reliable and real-time traffic flow prediction, optimize traffic scheduling, alleviate traffic pressure and improve the operation efficiency of the road network. The method is of great significance to the development of intelligent transportation.

The invention discloses a vibration signal-based rotating machinery fault direct intelligent diagnosis method. The method comprises the following steps of 1, sampling collected fault signals through an overlapping sampling layer; 2, segmenting samples through a convolutional layer; 3, training a sparse filter network subjected to L3 / 2 regularization through training data, obtaining a weight matrix, and extracting fault features of data segments through the weight matrix and a new activation function; 4, performing pooling dimension reduction on the features of the segments obtained by calculation through a pooling layer; 5, training an output layer, namely, a Softmax classifier by the pooled features, and classifying the extracted features of the samples; and 6, inputting the fault signalsto the trained network for performing diagnosis. The vibration signal-based rotating machinery fault direct intelligent diagnosis method realizes automatic extraction of the features and direct intelligent diagnosis of the fault signals, and is better in generalization capability and relatively high in accuracy and stability.