60 results about "Class prediction" patented technology

A method, system and a computer program product are provided for classifying elements in a ground truth training set by iteratively assigning machine-annotated training set elements to clusters which are analyzed to identify a prioritized cluster containing one or more elements which are frequently misclassified and display machine-annotated training set elements associated with the first prioritized cluster along with a warning that the first prioritized cluster contains one or more elements which are frequently misclassified to solicit verification or correction feedback from a human subject matter expert (SME) for inclusion in an accepted training set.

The invention provides an event detection model training method and device and an event detection method. The event detection model training method comprises the steps that training image frames in anumber of training videos with tags are acquired; a target neural network is used to extract feature vectors for the training image frame in each training video; in the unit of each training video, aself-attention mechanism processing network is used to carry out at least two rounds of weight assignment on a feature vector matrix formed by the feature vectors of each training video; the weight-assigned feature vector matrix is input into a class prediction network for class prediction to acquire a probability vector of the event classification result of the training video; and an event detection model is trained according to the probability vector of the event classification result and the result of the comparison between tag vectors formed by the tags of the training videos. According tothe method, computational resources and training time consumption are reduced without affecting the accuracy of the model.

The invention discloses a micro-blog emotion prediction method based on weak supervised type multi-modal deep learning and relates to the field of multi-modal emotion analysis. The method comprises the following steps of preprocessing micro-blog multi-modal data; carrying out the weak supervised training of a multi-modal deep learning model; and carrying out the micro-blog emotion prediction of the multi-modal deep learning model. The method solves the problems of the multi-modal discriminant expression and the data label limitation existing in the emotion prediction of the micro-blog multi-channel content in the prior art, and realizes the final multi-modal emotion class prediction, wherein the accuracy is adopted as the experiment evaluation standard. The consistency degree between the predicted micro-blog emotion polarity category and the pre-marked emotion category is reflected. The performance of the method is greatly improved and the correlation among multiple modals is considered. As a result, an optimal effect is achieved in the aspect of the overall multi-modal performance. An ideal classification effect is achieved for different emotion categories. Through the weak supervised training, an initial model for text and image modals is obviously improved in the aspect of emotion classification effect.

The embodiment of the invention provides a sample prediction method and device based on an intrusion detection model and an electronic device. The method comprises the steps of clustering samples in an initial training sample set, thereby obtaining first clusters, for each first cluster, if all samples in the first cluster belong to the same class, extracting the samples in the first cluster from the initial training sample set and marking the class of the first cluster as the class of any marked sample in the first cluster; obtaining a target training sample set, a target intrusion detection model and an initial test sample set, and for each sample in the initial training sample set, judging whether to extract the sample form the initial training sample set or not; obtaining a target test sample set composed of the samples which are not extracted from the initial training sample set; and carrying out class prediction on each sample in the target test sample set through utilization of the intrusion detection model. Through application of the scheme provided by the embodiment of the invention, when the sample prediction is carried out, the sample prediction accuracy is improved.

The present invention is a complete end-to-end video face recognition system. The invention performs a track-by-track labeling of all tracks within a video. A novel algorithm Mean Sequence SRC (MSSRC) is applied to a face track to perform video face recognition using a joint optimization to leverage all of the available video data and the knowledge that the face track frames belong to the same individual. Additionally the system constructs a probabilistic affinity graph combining appearance and co-occurrence similarities to model the relationship between face tracks in a video. Finally, using this relationship graph, random walk analysis is employed to propagate strong class predictions among similar face tracks, while dampening weak predictions.

The invention provides a system and method for predicting early development tendency of hot topics. The method includes the steps of collecting a topical time series; judging whether the series entersa recession period; if the sequence enters the recession period, adopting a clustering method as the complete topical time series to perform classification to obtain various topic classes, and substituting each topical time series of each topic class into a prediction model to perform training to obtain an intra-class prediction model for each topic class; if the sequence does not enter the recession period, analyzing similarity between a new topic time series and each complete topical time series in the topic classes, and taking average values as the matching degrees of new topics and the topic classes; screening out a set number of topic classes in descending order of the matching degrees; calling the intra-class prediction model of the topic classes screened out, and substituting the intra-class prediction model into the new topic time series to obtain a set number of predicted values; assigning different weight values to the predicted values to perform combination to obtain predicted values of new topics in the future time. According to the system and method, the early development tendency of a hot topic can be predicted accurately.

The present disclosure relates to a method and apparatus for predicting multi-label categories of images, an electronic device and a storage medium, the method comprising: performing feature extraction according to distillation features of an image to be predicted to obtain distillation feature information of the image to be predicted, wherein the distillation features comprising features obtainedafter knowledge distillation from a weak supervisory detection model used for class detection of the image; performing classification prediction processing on the distillation feature information toobtain the classification prediction confidence of the image to be predicted; determining the class prediction result of the image to be predicted according to the class prediction confidence of the image to be predicted. Embodiments of the present disclosure can realize distillation at a feature level, and can extract features from an image to be predicted using distillation features obtained after knowledge distillation from a weakly supervised detection model, so that a class prediction process of an image to be predicted is more efficient and a class prediction result is more accurate.

The invention provides a method for recognizing actions on the basis of deep feature extraction asynchronous fusion networks. The method is implemented by the aid of main contents including coarse-grained-to-fine-grained networks, asynchronous fusion networks and the deep feature extraction asynchronous fusion networks. The method includes procedures of inputting each short-term light stream stackof each space frame and each movement stream of input video appearance stream into the coarse-grained-to-fine-grained networks; integrating depth features of a plurality of action class grain sizes;creating accurate feature representation; inputting extracted features into the asynchronous fusion networks with different integrated time point information stream features; acquiring each action class prediction results; combining the different action prediction results with one another by the deep feature extraction asynchronous fusion networks; determining ultimate action class labels of inputvideo. The method has the advantages that deep-layer features can be extracted from the multiple action class grain sizes and can be integrated, accurate action representation can be obtained, complementary information in a plurality of pieces of information stream can be effectively utilized by means of asynchronous fusion, and the action recognition accuracy can be improved.

The invention discloses an optimized matting method and system capable of specifying a category. A semantic segmentation network of multi-class prediction is adopted to generate the ternary images ofdifferent class foregrounds of a to-be-tested image, a certain class of ternary images in the image can be rapidly obtained through the connected domain calculation, then an image matting network is adopted to carry out the refined image matting according to the selected class of ternary images and the to-be-tested image, so that an alpha image with optimized quality is obtained. According to thepresent invention, the foreground part of the specified category in the image can be automatically extracted, a transparent base map meeting various scene requirements is generated, the later image processing efficiency can be improved, the marketing personnel can also be helped to generate the marketing data, such as posters, etc., the commodity conversion rate is effectively improved, and the labor cost is reduced.

The invention discloses a federated learning data privacy protection method and system based on gradient perturbation, and belongs to the field of data privacy protection, and the method comprises the steps: carrying out the class prediction of a sample in a data participant through a local model after federated learning training, and obtaining an original prediction probability vector; disturbing the original prediction probability vector to obtain a disturbance prediction probability vector, wherein the prediction label of the disturbance prediction probability vector is the same as the prediction label of the original prediction probability vector, and the angular deviation of the gradient of the prediction loss function relative to the gradient of the prediction loss function of the original prediction probability vector is maximum; retraining each local model by taking the minimum difference between the original prediction probability vector and the disturbance prediction probability vector of each local model as a target; and aggregating the retrained local models to obtain a global model. The protected federal learning global model can effectively reduce the risk that the model prediction output and the model gradient leak the privacy of the user participants on the premise of maintaining the availability of the model.

The invention provides a medical data modeling-based class prediction method of necrotizing enterocolitis(NEC). The method includes the following steps: data sample selection, feature screening, important feature selection, model establishment and prediction result scoring. According to the method, systematic analysis and modeling are carried out on relevant data of the necrotizing enterocolitis, an evaluation method of model prediction is given, through a model, effective auxiliary diagnosis can be carried out on the necrotizing enterocolitis of patients based on the data of the necrotizing enterocolitis, effective prevention, intervention and treatment are enabled to be carried out at an initial disease period, and a basis is provided for achieving the best treatment effect.

The invention provides a detection model training method and device, a target detection method and device and an electronic system, and relates to the technical field of target detection, and the detection model training method and device comprises the steps: obtaining an image sample corresponding to a current period from an image sample set in at least one training iteration period, obtaining a feature map of each image sample through a backbone network, and then generating a new sample label based on the feature maps of the image samples and the original labels on the image samples; and training a classification network, a regression network and a multi-class prediction network based on the image samples after the new sample labels are distributed, controlling the number of training iterations by taking loss functions corresponding to the three networks as training constraint conditions, and establishing a detection model for target detection until training is stopped. Through the invention, the problem that a detection model has relatively limited target labeling information in the image can be relieved, and the training precision of the detection model is improved.

The invention provides an SAR automatic object recognition method based on heterogeneous CNN (Convolutional Neural Network) integration, and relates to the field of radar target recognition. The method employs a heterogeneous CNN, achieves the full extraction of the target feature information in an SAR image, and achieves the target class prediction. Based on the integrated learning theory, the method integrates the results of a plurality of CNNs, obtains a final class label, and achieves the quick and precise recognition of an SAR target. According to the invention, the method can achieve theadjustment of the number of heterogeneous networks according to the actual hardware conditions and a performance index, still can guarantee the higher recognition rate under the condition that the number of training samples is smaller, is high in efficiency, is stable, and is strong in generalization capability.

The invention discloses a bus punctuality prediction method based on GPS data, and belongs to the field of public transport information processing. The method comprises the steps: the collection and processing of bus GPS data and departure data, the determining of a bus punctuality value, the extraction of bus punctuality influence factors and the prediction of bus punctuality. The method employs an SVM (support vector machine) algorithm. The method combines the bus GPS data, extracts the track information and departure information of a plurality of buses, is convenient and quick, and reduces the data processing cost. Moreover, the method employs the SVM algorithm for the two-class prediction of the punctual conditions of a downstream stop, and enables passengers to know the bus operation conditions better to reasonably adjust the travel time. Meanwhile, the method enables a bus operation department to be able to timely adjust the departure interval of buses, and improves the bus service level.

An embodiment of the invention discloses a fabric flatness objective evaluation method and a fabric flatness objective evaluation device on unsupervised machine learning, wherein the method comprises the steps of acquiring sample data in a standard evaluation environment; preprocessing the acquired sample data, eliminating background information and interference information of an image; vectorizing the preprocessed data by means of computer image processing technology; classifying the vectorized data, and generating a characteristic reference set; and performing image class prediction on the characteristic reference set, thereby obtaining an evaluation result. In the fabric flatness objective evaluation method and the fabric flatness objective evaluation device, through extracting and abstracting a bottom-layer characteristic, a fabric image is vectorized; clustering is performed according to the characteristic of the fabric image, and a label is set for a clustering result. Through unified extraction and abstraction on the bottom layer characteristic and objective reference classification, fabric grade prediction is performed, thereby obtaining an evaluation result in a more fair and objective manner, reducing an error caused by artificial adoption of data for training, and furthermore preventing a subjective error caused by artificial evaluation.

The invention discloses a vehicle real-time detection method based on a micro-convolution neural network, which comprises the following steps: (1) preprocessing an input image, converting the input image into a gray-scale image, and normalizing the gray-scale value of the gray-scale image to be between [0, 1] or [1, 1] and reassembled to a uniform size; (2) inputting the image data obtained in thestep (1) into a 7-layer micro-convolution neural network, training the micro-convolution neural network, and generating prediction boxes of different scales for class prediction and regression targetposition; (3) training records the error on the training set and the test error on the verification set of each iteration; 4) judging whether that los on the successive 5 iterative verification setsis reduced, if so, returning to the step 2) if the loss is not reduced, terminating the training, saving the parameters of the 7-layer microconvolution neural network, and checking the feature extraction effect. The invention uses 7-layer convolution neural network structure instead of complex VGG (Deep Convolution Neural Network for Large Scale Image Recognition), which can be trained and testedon ordinary machines, does not need high performance computing equipment such as GPU (Graphics Processor) with super performance, nor does it need pre-trained network, it can be trained and tested from scratch.

The invention discloses a text classification method based on part-of-speech classification. The method comprises the steps of acquiring a training text set and a test text set from a network; preprocessing the training text set and the test text set to obtain a plurality of word sets of each text in the training text set and the test text set; training a text topic generation model LDA by takingthe obtained word sets of each text as inputs to obtain a text-word set-topic mixed probability distribution model of each text under different topic numbers; by using an SVM-train function, carryingout classifier training on a plurality of text-word set-topic mixed probability distribution models to obtain a plurality of trained classifiers; and by using the text-word set-topic mixed probabilitydistribution models as inputs of the trained classifiers, carrying out SVM class prediction. The method can solve the technical problems of high dimension of needed characteristic words during modeltraining, low classification accuracy and poor generalization capability of the classifiers in an existing method.

A device for assigning one of a plurality of predetermined classes to each pixel of an image, the device is configured to receive an image captured by a camera, the image comprising a plurality of pixels; use an encoder convolutional neural network to generate probability values for each pixel, each probability value indicating the probability that the respective pixel is associated with one of the plurality of predetermined classes; generate for each pixel a class prediction value from the probability values, the class prediction value predicting the class of the plurality of predetermined classes the respective pixel is associated with; use an edge detection algorithm to predict boundaries between objects shown in the image, the class prediction values of the pixels being used as input values of the edge detection algorithm; and assign a label of one of the plurality of predetermined classes to each pixel of the image.

The invention provides a three-dimensional face recognition method based on sparse features. The method comprises the steps of: S1, acquiring three-dimensional face information and pre-processing thethree-dimensional face information; S2, extracting sparse features with classification information through sparse representation and dictionary learning of the three-dimensional face information; S3,performing class prediction on the extracted sparse features by using a classifier to obtain a final classification result. The three-dimensional face recognition method based on sparse features has the advantages that the K-SVD algorithm integrating sparse representation and dictionary learning is used for feature extraction of three-dimensional face models, so that the data dimensionality and the calculation cost are reduced; the algorithm is simple and easy to implement and has higher recognition rate and greater robustness.

A method and apparatus for predicting a category of information is disclosed in embodiments of that present application. One embodiment of the method includes: receiving a prediction request of a userfor class prediction of information to be classified; obtaining, based on the prediction request, a set of information to be classified and a target screening class information group, wherein, the screening class information group is associated with a class label; determining whether the information to be classified satisfies each piece of filtering class information in the filtering class information group for the filtering class information group in the target filtering class information group set, and if so, determining a class label associated with the filtering class information group asa target class label corresponding to the information to be classified; and generating correspondence relation information for characterizing a correspondence relationship between the information tobe classified and the target class label. By utilizing the set of target screening class information sets, this embodiment can realize information class prediction even if the machine learning model used for classification does not cover it.

The invention discloses a protein structure class prediction method based on weighted composition position vectors and a support vector machine. Firstly, protein structure class normal data sets are selected or established, weighting factors are introduced based on composition position vectors, and to-be-predicted protein sequences are represented through the weighted composition position vectors; the weighted composition position vectors are combined with the support vector machine, and the protein structure class prediction method is established through a direct multiclass classification strategy. According to the method, information of amino acid compositions is contained, position information of each amino acid residue in the protein sequences is also contained, the one-to-one correspondence function relation is formed between the position information and the protein sequences, the weighting factors are introduced to the composition position vector representation method, and prediction accuracy can be obviously improved by means of adjusting the weighting factors. The method is simple, rapid and sensitive, and is expected to be applied to other protein prediction fields.

The invention provides an information acquisition method and system based on social media emergencies. The method comprises the following steps: S1, constructing a corpus of emergencies; S2, using a support vector machine classifier to carry out non-emergency event classification filtering, and achieving first-stage classification; And S3, performing positive and negative class prediction classification by using a naive Bayes classifier to realize second-stage classification. According to the invention, corpus acquisition of related keywords is carried out on the social media through the crawler; A support vector machine classifier is used for carrying out non-emergency event classification filtering, first-stage classification is achieved, a naive Bayes classifier is used for carrying outpositive and negative class prediction classification, second-stage classification is achieved, the information classification precision is improved by 2.9% compared with the result of non-instant seismic information screening, and the accuracy of information classification is improved by 2.9%. Compared with the prior art, the method has the advantages that the value of the F-Masure is increasedby 2.6%, the problem that the text classification result is low in precision in the prior art is solved, the classification precision is improved, a decision maker can control disaster events easily,and a basis is provided for decision making.

The invention discloses a deep learning model security vulnerability testing and repairing method, device and system based on a genetic algorithm. An evaluation function is constructed according to the class prediction value difference of a label of an input image in a deep learning model and the added noise minimization, the evaluation function is optimized to generate a malignant image which can cover the boundary of the deep learning model and has a large identification gap, when the malignant image can be obtained, it shows that the original deep learning model is unsafe, otherwise, the original deep learning model is safe, so that the security test of the deep learning model is realized, and the security test method is simple and accurate; and the obtained test image is used to carry out intensified training on the original deep learning model so as to repair the deep learning model and improve the recognition result accuracy of the deep learning model.

The invention discloses a data processing method and device, equipment and a readable storage medium, and the method comprises the steps: obtaining a target image, and identifying N prediction regions of a target object in the target image; obtaining a prediction object category and a prediction object label feature corresponding to the target object in each prediction area; obtaining the maximum class prediction probability from the class prediction probabilities corresponding to the N prediction object classes, and determining the prediction region corresponding to the prediction object class with the maximum class prediction probability as a target prediction region; determining a coverage area commonly covered by the residual prediction area and the target prediction area; determining the label feature similarity between the prediction object label features corresponding to the remaining prediction regions and the prediction object label features corresponding to the target prediction region; and determining an optimal prediction region of the target object in the target image according to the similarity between the coverage region and the label feature. According to the invention, the identification accuracy of the object in the image can be improved.