165results about How to "Efficient training process" patented technology

The invention belongs to the field of face recognition, in particular to a neural-network-based large-scale unbalanced data face recognition method and system, and aims to solve the problems of large-scale data optimization and improvement of face recognition efficiency. According to the method, the performance of model face recognition is improved by improving the loss function and the sampling mode, the loss function of the self-adaptive boundary margin is provided in the aspect of the loss function to cope with unbalanced face data, and an improvement scheme is provided in the aspect of sampling for data sampling and classification template sampling. According to the method, model training can be efficiently carried out on large-scale unbalanced face data, and the performance is improved.

The invention discloses a training method and device for a financial risk identification model, computer equipment and a medium, and the method comprises the steps: obtaining first user data, with a first credit risk label, of a target domain financial project, inputting the first user data into a meta-learning device for training, and obtaining a classifier corresponding to the target domain financial project for risk identification. According to the method, the classifier corresponding to the target domain financial project is trained in a meta-learning mode, priori knowledge in a source domain task can be effectively migrated, so that the data volume of labeled samples required by model training is small, the generalization performance of the recognition model is improved, and the training process of the model is faster and more efficient. Besides, in the training process of the meta-learner, the source domain correlation among the categories of the task sets is learned, so that prior knowledge can be effectively migrated from tasks closer to the current target domain task during migration learning, and the accuracy of model recognition can be improved. The method can be widelyapplied to the technical field of machine learning.

The invention provides a federated learning model training method based on data feature perception aggregation. According to the method, a cloud server calculates a global optimal model according to alocal model uploaded by an edge client, performs dimension reduction on data characteristics according to mined edge client data characteristics, screens out key characteristics, and performs clustering grouping on the edge client based on the key characteristics; the cloud server issues grouping information and a global optimal model to the edge client; and the edge client provides local data according to the received global optimal model and provides local execution model training according to the global optimal model, and one local model is randomly selected from the same edge client groupor the optimal model in the edge client group is selected and uploaded to the cloud server. According to the method, the data characteristics of the edge clients are fully utilized, the edge clientsare grouped, unnecessary communication is avoided, the influence of statistical isomerism on model training is relieved to a great extent, and the training efficiency of the model is improved.

The invention provides a visual question-answering model training method and device, and relates to the technical field of computers. The visual question-answering model training method comprises thesteps of obtaining a training sample and a sample label; extracting sample image feature information and sample problem feature information; performing feature cross processing on the sample image feature information and the sample problem feature information to obtain a sample image feature vector carrying the sample problem information and a sample problem feature vector carrying the sample image information; inputting the sample image feature vector carrying the sample question information and the sample question feature vector carrying the sample image information into the visual question-answering model to obtain a prediction answer through the visual question-answering model; determining a loss value of a loss function based on the real answer and the predicted answer; and updating the visual question and answer model through the loss value of the loss function.

The invention relates to a target detection method for an SAR (Synthetic Aperture Radar) image based on deep reinforcement learning, which comprises the steps of S1, setting the number of iterations,and sequentially processing images in a training set in each iteration process; S2, inputting an image from the training set, and generating training samples by using the Markov decision process; S3,randomly selecting a certain number of samples, training the Q-network by adopting a gradient descent method, obtaining the state of a reduced observation area, generating a next sample until a presettermination condition is met, and terminating the image processing process; S4, returning back to the step S2, continuing to input the next image from the training set until all images are completelyprocessed, and terminating the current iteration process; S5, continuing the next iteration process until the set number of iterations is met, and determining network parameters of the Q-network; andS6, performing target detection on an image in a test set through the trained Q-network, and outputting a detection result. The target detection method obtains good detection accuracy in target detection for the SAR image.

The invention discloses a rotary machine residual life prediction method of a multi-layer bidirectional gating cycle unit network. The rotary machine residual life prediction method comprises the following steps: acquiring a vibration signal; constructing health indexes; constructing a network training set; constructing a multi-layer bidirectional gating circulation unit network; training a multi-layer bidirectional gating cycle unit network; and carrying out network testing, residual life estimation and confidence interval acquisition; and performing residual life prediction evaluation. According to the rotary machine residual life prediction method, the advantages of strong feature extraction capability of deep learning are combined, and regression prediction is carried out by utilizinga bidirectional gated cycle unit neural network, and the confidence interval of the residual life is obtained through a Bootstrap method. Aiming at the problems that the model precision is sensitive to the value of the learning rate in the training process of a recurrent neural network model, the prediction performance of the model is influenced by too high and too low values, and the neural network is efficiently trained by utilizing the natural exponential decay learning efficiency. The rotary machine residual life prediction method can accurately predict the residual life and confidence interval of the rotary machine, can greatly reduce the expensive unplanned maintenance, and avoids the occurrence of a big disaster.

The invention discloses a convolutional echo state network time sequence classification method based on a multi-head self-attention mechanism. The method is realized based on two components of an encoder and a decoder. The encoder is composed of an echo state network of a multi-head self-attention mechanism. An input time sequence is firstly subjected to high-dimensional mapping coding through a plurality of echo state networks to generate original echo state characteristics, and then re-coding of echo state characteristic global space-time information is achieved based on a self-attention mechanism, so that the re-integrated high-dimensional characteristic representation has higher discrimination capability; And finally, taking a shallow convolutional neural network as a decoder to realize high-precision classification. The classification model established by the method inherits the characteristic of high training efficiency of the echo state network, realizes the re-coding of the space-time characteristic information by introducing a self-attention mechanism model without additional parameters, achieves a high-precision time sequence classification effect, and is a simple and efficient model.

The invention provides a classification model application and a classification model training method and device, and the method comprises the steps: obtaining a to-be-processed corpus, and convertingthe to-be-processed corpus into a word unit sequence; determining a first occurrence frequency of each keyword in a pre-constructed keyword set in the word unit sequence, and generating a first feature vector based on the first occurrence frequency corresponding to each keyword; determining a second occurrence frequency of each word in the word unit sequence in the word unit sequence, and generating a second feature vector based on the second occurrence frequency of each word and a preset inverse document frequency of each word; inputting the first feature vector into a first classification model, and outputting a first classification result of the to-be-processed corpus; and inputting the second feature vector into a second classification model, and outputting a second classification result of the to-be-processed corpus, and determining the category of the to-be-processed corpus based on the first classification result and the second classification result. By means of the method, thecorpus classification accuracy can be improved.

The present invention discloses an image super-resolution method based on active sampling and Gaussian process regression, which mainly aims to solve the problem of poor super-resolution effects of a texture region in the prior art. According to the method, a training set is constructed by fully using a large number of existing external natural images, and a sample information amount is measured by using characterization and diversity as indexes, so as to extract a simplified training subset, which enables training of a Gaussian process regression model to be more efficient. The implementation steps are as follows: 1. generating an external training sample set, and performing active sampling on the external training sample set to obtain a training subset; 2. learning a Gaussian process regression model based on the training subset; 3. preprocessing a test image and generating a test sample set; and 5. applying the learned Gaussian process regression model on the test sample set, and predicting and outputting a super-resolution image. The method has a relatively strong super-resolution ability and is capable of restoring more detail information in regions such as texture, so that the method can be used for video monitoring, criminal investigation, aerospace, high definition entertainment and video or image compression.

The invention discloses a rhythm predicting method and a system wherein the method comprises: creating a text rhythm model in advance; gathering text data with corresponding voice data; based on the rhythm information of the voice data, performing automatic rhythm marking to the corresponding text data; obtaining the automatically marked text data; using the automatically marked text data to train the text rhythm model; receiving the to-be-predicted text data; then, extracting the text characteristics of the to-be-predicted text data; and finally utilizing the text characteristics and the text rhythm model to perform rhythm predictions to the to-be-predicted text data. According to the method and the system, as the gathered text data all correspond to voice data and the voice data contain the rhythm information in reality, it is possible to perform automatic rhythm marking to the text data. Therefore, with the system and the method, the problem in the prior art can be solved that artificial marking is required for the rhythm edges of all trained text data to train the text rhythm model, which leads to the high human cost and the long consumption of time.

The invention provides a welding image defect recognition method and device, a storage medium and equipment. The method comprises the steps of acquiring a to-be-recognized welding image; inputting theto-be-recognized welding image into a defect recognition model, recognizing the welding image through the defect recognition model to obtain a welding image type, and judging whether the welding image has defects or not and the defect type when the defects exist or not; wherein training processing is carried out on the initial defect recognition model as follows: forming a defect recognition model by combining a convolution basis of a pre-training model with a full-connection classifier; carrying out migration training on the defect recognition model; wherein migration training refers to freezing a convolution basis, training a full-connection classifier, and then finely adjusting a defect recognition model. Based on the combination of deep learning and transfer learning technologies, efficient training of the defect recognition model can be realized under the condition of limited sample number, and the accuracy of welding image defect recognition is improved.

The invention relates to a high-voltage line insulator defect detection method based on a generative adversarial network. The method comprises the following steps: establishing an image data set; constructing an improved GANomaly model, wherein the improved GANomaly model sends a feature map output by each convolution layer in the encoder sub-network to the decoder sub-network for feature fusion;constructing a loss function of the improved GANomaly model; training an improved GANomaly model; sending the defective insulator image in the test set into a trained improved GANomaly model to obtainan adaptive threshold; sending a to-be-detected insulator image to the trained improved GANomaly model to judge the image type. Compared with the prior art, the method solves the problems that the number of negative samples in the insulator image is small, the negative samples are difficult to collect, a large amount of energy is needed to label data, and the feasibility and adaptability of the training process are improved.

The invention relates to a machine learning algorithm whole-process training method and system based on a cloud infrastructure. The method comprises the steps that a training data set and a to-be-solved problem data set are uploaded to a cloud data server through a web application program, and a preset machine learning algorithm is selected; the training data set is obtained by a cloud computing server through a cloud data server, and model training is conducted by using the training data set to obtain a training model; the to-be-solved problem data set is obtained by the cloud computing server through the cloud data server, solving results are obtained by using the training model according to the to-be-solved problem data set, and the solving results are returned to the cloud data server;the solving results are returned to the web application program by the cloud data server. According to the method, computing environments and computing resources are deployed at cloud ends to help technical researching and developing workers in efficiently and rapidly constructing a machine learning algorithm model and efficiently training the model, and therefore time cost and software and hardware purchasing cost are reduced.

The invention discloses a weather image intelligent identification method and system based on multi-depth convolutional neural network fusion, and the method comprises the following steps: A, collecting an image, and carrying out the preprocessing of the image; B, establishing four deep convolutional neural network models with different structures, respectively improving full connection layers ofthe deep convolutional neural network models, adding a feature layer, and performing network training based on the high-quality training data obtained in the step A; C, extracting features of newly added feature layers of the four obtained deep learning models, and performing training by adopting an Xgboost ensemble learning model to obtain a fusion model; and D, amplifying a weather image to be identified, identifying the amplified weather image by the obtained fusion model, and performing voting to obtain an identification type with the highest final vote number. According to the method, theproblem of low efficiency of manually extracting the image features in an existing traditional weather recognition method is solved, and the recognition accuracy of the deep learning model is effectively improved.

The invention belongs to the field of prediction of amorphous alloy thermoplasticity forming performance, and discloses an artificial neural network predicting method of the amorphous alloy thermoplasticity forming performance. The predicting method comprises the following steps of a, selecting multiple performance parameters and collecting data of the performance parameters, dividing the data into a training sample, a verification sample and a to-be-predicted sample, and testing to obtain feature index test values corresponding to the training sample and the verification sample; b, selectingan artificial neural network model as an initial predicting model for the amorphous alloy thermoplasticity forming performance, adopting the training sample to train the artificial neural network model, and determining an improved predicting model; c, adopting the verification sample to verify the improved predicting model, finally obtaining a final predicting model, and adopting the final predicting model for prediction. By means of the artificial neural network predicting method of the amorphous alloy thermoplasticity forming performance, the amorphous alloy thermoplasticity forming performance is effectively predicted without experiments, guidance is provided for development of an amorphous alloy system suitable for thermoplasticity forming, the time for developing the new amorphous alloy system is greatly shortened, and the money cost for developing of the new amorphous alloy system is greatly reduced.

The embodiment of the invention discloses a clustering model training method and device, electronic equipment and a computer storage medium. The method comprises the steps that a global clustering model is acquired based on master nodes; according to any one of at least one slave node in a distributed system, the global clustering model is acquired from the master nodes in the distributed system,clustering estimation is performed based on the global clustering model and calculation data allocated to any corresponding slave node, and a local clustering model corresponding to any slave node isobtained; the distributed system comprises the master nodes and at least one slave node in communicating connection with the master nodes, wherein the master nodes are in communicating connection withall the slave nodes; and the global clustering model is trained based on the obtained local clustering models corresponding to all the slave nodes. Through the clustering model training method in theembodiment, the synchronization rate among the calculation nodes is lowered, and clustering efficiency is improved.

The invention discloses a moving vehicle detection method based on aerial video of an unmanned aerial vehicle, at first, the SURF algorithm is used to match the feature points of the image and removethe anomaly point, a UAV image registration algorithm combining global and local homography matrices is used to obtain the transformation matrix, compensating for the adverse effects of airborne camera motion, a 2-frame difference method is used to reduce the area to be detected, and then traverses the region to be detected according to the center of the superpixel, then, the low-order features are extracted by using the multi-channel HOG feature algorithm, and the high-order features are obtained by introducing the context information of the vehicle, and the two features are fused to obtain the multi-order features of the target vehicle. Finally, the multi-order features and dictionary learning algorithm are combined to realize the moving vehicle detection. The method can suppress the influence of the motion of the airborne camera of the unmanned aerial vehicle, process the vehicle deformation and background interference in the image, and improve the robustness and real-time performance of the moving vehicle detection.

The present invention discloses a word vector generation method based on the Gaussian distribution. The method comprises: firstly, preprocessing the corpus; secondly, using the punctuation to performtext division on the corpus; then combining the local and global information to infer the word meaning, and determining the mapping relationship between the word and the word meaning; and finally, obtaining a word vector by optimizing the objective function. The innovations and beneficial effects of the technical scheme of the present invention are as follows that: 1, words are represented based on the Gaussian distribution, point estimation characteristics of traditional word vectors are avoided, and more abundant information such as probabilistic quality, meaning connotation, an inclusion relationship, and the like can be brought to the word vectors; 2, multiple Gaussian distributions are used to represent the words, so that the linguistic characteristics of a word in the natural language can be coopered with; and 3, the similarity between the Gaussian distributions is defined based on the Hellinger distance, and by combining parameter updating and word meaning discrimination, the number of word meanings can be inferred adaptively, and the problem that the number of hypothetical word meanings of the model in the prior art is fixed is solved.

The invention relates to a network intrusion detection method combining longicorn stigma with a random forest, which is used for processing a network intrusion monitoring problem based on machine learning and can be used for training a network intrusion monitoring model with high accuracy by using less time. When abnormal flow passes, the detection system can find a problem through flow analysis and generate a corresponding signal. Compared with a particle swarm optimization random forest algorithm and a fruit fly optimization random forest algorithm, the method can complete model training anddetection more efficiently.

The invention discloses watermark recognition online training, sample preparation and removal methods, a watermark recognition online training device, equipment and a medium. The training method comprises the steps: extracting a background image from a background data set, and randomly cutting the background image into a preset specification; extracting a watermark original image from the watermark data set, and performing image deformation processing on the watermark original image to form a watermark enhanced image; synthesizing the watermark enhancement image and the background image to obtain a corresponding synthesized image; and feeding the composite image as a training sample into a preset image segmentation model which has a coding and decoding structure and is suitable for capturing multi-scale features to implement sample training, so that the image segmentation model is suitable for removing the watermark original image from the to-be-identified image through the training. According to the invention, the watermark-containing training sample can be automatically synthesized on line and immediately used for training the image segmentation model, so that the model has the capability of identifying and removing the watermark in the picture.