68results about How to "Classification results are reliable" patented technology

The invention provides a diabetic retinopathy grade classification method based on deep learning. The diabetic retinopathy grade classification method comprises the steps of: constructing a sample library; removing backgrounds and noise of ophthalmoscope photographs in the sample library; normalizing the images of different brightness and different intensity to the same range by adopting a local mean value subtracting method; adopting random stretching and rotating methods for different samples for data augmentation, and constructing a training set and a test set; training an initial deep learning network model by establishing an input portion architecture, a multi-branch feature transformation portion architecture and an output portion architecture separately; and inputting samples to betested into the trained initial deep learning network model for diabetic retinopathy grade classification. Compared with the traditional processing method, the diabetic retinopathy grade classification method gets rid of the dependence on prior knowledge, and has good generalization ability; and by adopting the designed multiple grades, a small-sized convolution kernel can be used for extracting very tiny lesion features, thereby making the classification results more reliable.

The invention provides a deep learning based video crowd emotion analysis method, and mainly relates to classification for crowd emotions in the video by using a multi-stream neural network. The method comprises the steps of building a multi-stream neural network (pixel, optical flow and saliency), concurrently extracting features in pixel information, superposition optical flow information and saliency information of a video sequence by using the network, and finally integrating the three types of features to obtain classification for the crowd emotions. The self-learning ability of deep learning is given into full play, the limitation of manual feature extraction is avoided, and the method is enabled to be higher in adaptability. Training and prediction are performed by using structural features of the multi-stream deep learning network, classification results of multi-stream sub-networks are integrated finally, and the accuracy and the work efficiency are improved.

The invention relates to an intelligent biological signal anomaly detection method and a system capable of carrying out efficient, accurate and low-cost automatic analysis on biological signals. The system comprises a model training stage and a model application stage, wherein the model training stage comprises the following steps: acquiring a biological signal; preprocessing the acquired signal;sending the preprocessed data to a CNN-BilSTM-Attention combination model to extract features, wherein the signal firstly enters a CNN convolution neural network to extract the spatial features, thenenters a two-way long-short time memory network to extract time features, then passes through each network connection layer and finally passes through a SoftMax layer to calculate the final result, each link of the network is established, and the emphasis degree of different sequence points of signals are optimized by setting an Attention mechanism; calculating loss of the network module until themodule loss is less than the threshold, then the model training is completed, and otherwise Adam is used to optimize model parameters. The invention effectively improves the analysis effect of biomedical signals.

The invention provides a face sequence expression analysis method based on deep learning and mainly relates to classification of face sequence expressions through a multi-scale face expression recognition network. The face sequence expression analysis method comprises the steps that the multi-scale face expression recognition network (comprising three channels for processing face sequences with different resolution ratios of 128*128, 224*224, 336*336 and the like) is constructed, features in the face sequences with the different resolution ratios are parallelly extracted through the network, finally, the three kinds of features are fused, and thus classification of the face sequence expressions is obtained. According to the face sequence expression analysis method based on deep learning, the self-learning capability of deep learning is given to full play, limitation of manual feature extraction is avoided, and thus the adaptability of the method is higher; and parallel training and predicting are conducted through the structural features of a multi-stream deep learning network, finally, the classification results of a plurality of sub-networks are fused, the accuracy rate is increased, and the working efficiency is improved.

The invention discloses an unmanned aerial vehicle airborne platform-based vehicle type identification method and belongs to the technical field of video images. The method comprises the following steps of S1: adjusting the height of an unmanned aerial vehicle above the ground to a position suitable for vehicle type identification; S2: acquiring infrared image positive and negative samples used for vehicle target detection offline, performing characteristic extraction by utilizing a convolutional neural network (CNN), and performing support vector machine (SVM) training to obtain an SVM classifier model used for online vehicle type identification; S3: performing infrared photography by utilizing an infrared video camera to obtain a video image signal; and S4: performing sliding window sampling on a shot infrared image sequence, extracting vehicle type characteristics by utilizing the CNN, and inputting the characteristics in the classifier model obtained in the step S2 to perform classification. According to the method, the vehicle type identification can be finished in the day, at night and under the condition of severe weather condition and relatively low visibility by adopting the infrared video camera; and meanwhile, the method can be used for dynamically selecting a detection region, so that the flexibility of a detection system is improved.

A valve short circuit fault classification and positioning method of a high-voltage direct current transmission converter comprises the steps: a, if a high-voltage direct current transmission protection system determines that a converter has valve short circuit faults, the processor of a protection system reads current magnitude information collected by a current measurement device in a sampling window; b, the processor performs integration calculation of section electric charge quantity and branch electric charge quantity differences; c, the processor distinguishes the alternating current side short of the converter, the bridge arm short circuit and the direct current outgoing line short circuit according to the size relation of each section electric charge quantity; and d, if the bridge arm short circuit happens, the processor further locates a fault bridge arm according the sizes of the branch electric charge quantity differences and the direction of the alternating current branch current. The method provided by the invention can rapidly classify positioning valve short circuit faults to facilitate rapidly and effectively processing faults and prevent accidents from being enlarged; and the method provided by the invention is not influenced by fault moments, noise and the length of the sampling window, and is low in sampling rate requirement and good in adaptability.

The invention provides a method and a system for automatically classifying objects. The method comprises the following steps: inputting a group of the objects and a class hierarchy for classifying theobjects; acquiring body information correlative with a field; using a method based on the body information to classify each object into the class hierarchy so as to obtain a first classification result; using a method based on information retrieval (IR) to classify each object into the class hierarchy so as to obtain a second classification result; and combining the first and the second classification results to obtain a final classification result. The method and the system can be applied to single-layer or multi-layer class hierarchies, and can effectively and accurately achieve object classification so as to obtain a smaller classification result granularity.

The invention provides a target feature-assisted multi-source data correlation method, aiming to provide a correlation method with high utilization rate of measurement parameters and capable of improving the correlation accuracy of a radar and electronic support measures (ESM). The target feature-assisted multi-source data correlation method is realized through the following technical scheme thataccording to the correlation between heterogeneous features, a correlation classification rule of heterogeneous sensor data is determined, a mapping correlation model of a target motion feature space,a target recognition feature space and a target type space is established, a category identification frame is established, K neighbors being at a distance from target features are found according toa K-K-nearest neighbor-NN rule, and trust assignment is constructed based on the distance between a target and the neighbors of the target, an acceptance threshold value and a rejection threshold value; the features of the target at each sampling time are obtained, and then BK-NN training is carried out on the target features at each t time, local static evidences at corresponding times of the categories are obtained and are integrated to generate a static criterion; and the comprehensive results of dynamic classification of different features are calculated, and the correlation filtering results are obtained.

The invention relates to the field of banknote anti-counterfeit research, and specifically relates to a banknote authentication method based on a random forest. A multispectral image of a banknote acquired based on a sensor is processed using an embedded equipment chip. The method comprises the following steps: carrying out feature extraction in a specific area cut out based on an acquired original image and a gradient image thereof to get multiple features needed for classification; creating a single decision tree, and carrying out node splitting from a root node to leaf nodes; and building a random forest including multiple decision trees. Through the scheme, a random forest classifier which can be realized on embedded equipment is established, banknote authentication can be performed on a currency counter using a DSP chip as a processor, the banknote authentication of the currency counter is improved, counterfeit banknotes mixed in real ones can be accurately identified and classified, computation is quick, and the real-time requirement of the equipment is satisfied. The defect that the traditional authentication method has weak ability to identify unknown counterfeit banknotes and new currencies is overcome.

The invention discloses a news theme classification method which is characterized by including the following steps: (1) building a seed dictionary according to news theme classes, conducting word segmentation on news themes to extract title keywords, conducting meta searching on the title keywords through a plurality of search engine servers based on the internet, conducting frequency statistics on seed keywords in meta searching results, and judging the final news theme classes according to the occurring frequency of the seed keywords in the meta searching results. According to the news theme classification method, the classification time can be greatly shortened, the labor cost is effectively reduced, dependency on historical data is avoided, the use time of the whole classification process is short, the classification result is reliable, multiple kinds of classifications can be carried out on news, and the news theme classification method is high in universality in actual situations.

The invention provides a BP neural network-based micro-energy device energy identification method, which comprises the following steps: S1, collecting a dynamic voltage of a micro-energy device in anopen-circuit state to obtain an original voltage signal, and performing adaptive threshold wavelet transform processing on the original voltage signal to remove noise; S2, extracting an R wave peak value of the denoised voltage signal so as to obtain model input data; S3, establishing a BP neural network model, inputting data to train the model, and stopping training when a training error is smaller than a preset value to obtain a qualified BP neural network model; and S4, identifying a to-be-identified voltage signal by using the BP neural network model obtained in the step S3. According to the invention, accurate and rapid energy identification and classification can be carried out, and the classification result is reliable. The method is high in anti-interference capability. A pluralityof characteristic quantities with relatively high influence proportions in energy signal comparison of the micro-energy device are selected.

The invention relates to an intelligent LED (light emitting diode) display screen classifying system and method. The system comprises a database, an analysis algorithm module and a classifying module, wherein the database is used for storing LED display screen relevant data including operating state parameters of LED display screen and user operation data; the analysis algorithm module is provided with a plurality of algorithm sub-modules which are respectively used for analyzing, creating sorting of the corresponding data acquired from the database and for screening a sorting result through a threshold value so as to generate LED display screen classifying result data to be outputted; the classifying module is provided with a plurality of categories, each category corresponds to the corresponding algorithm sub-module in a plurality of algorithm sub-modules in the analysis algorithm module, and the classifying module is used for responding to the selection operation for a plurality of categories and transmitting the corresponding classifying request to the analysis algorithm module so as to trigger the corresponding algorithm sub-module in a plurality of algorithm sub-modules. The categories are defined as collections with evidence and real significance, the LED display screens are classified according to objective data, and the classifying result is realistic and credible.

The invention discloses a reservoir classification evaluation method based on secondary parameter selection. The reservoir classification evaluation method comprises the steps of introduction and classification of reservoir evaluation parameters, correlation analysis of the reservoir evaluation parameters, primary selection of the reservoir evaluation parameters, secondary selection of the reservoir evaluation parameters and determination of reservoir classification evaluation standards through K-means clustering. Beneficial effects of the present invention are: reservoir evaluation parametersare determined through secondary optimization; reservoir evaluation is completed through K-means clustering; compared with the prior art, the reservoir classification evaluation method has the advantages that parameter selection is more accurate and reliable, the reservoir classification result is more credible, reservoir evaluation parameters can be quantitatively optimized through secondary selection of the reservoir parameters, the standard is accurate and unified, and the reservoir evaluation method combines macroscopic and microscopic parameters, so that the finally obtained evaluation standard is more comprehensive and accurate.

The invention provides a passive terahertz security check method and system and a medium, and the method comprises the steps: judging whether a collected terahertz image is a first strip-shaped background image or not, and obtaining a global mean line vector of the image; removing stripe interference information in the original image; performing gray histogram extraction; distinguishing the difference between the background and the foreground by adopting an image segmentation algorithm, and extracting a segmentation threshold; performing secondary threshold segmentation processing according tothe segmentation threshold, and removing background interference; performing histogram equalization operation on the pixels; performing morphological processing; performing weighting processing on the image to obtain a final image; and establishing a YOLOV3 framework based on deep learning, configuring an operation environment, taking a final image as input for detection, and outputting a final type of suspicious dangerous goods carried by a human body. The deep learning detection has a higher recognition rate and a more reliable classification result, and successfully solves the problem thathidden suspicious dangerous goods are difficult to accurately recognize in a terahertz image.

The invention discloses a network security blog classification method and a network security blog classification system based on feature extraction. The method comprise calculating the non-dictionaryword density for each blog; calculating blog malicious tendencies for each blog; counting high-frequency words common to all blogs; calculating the frequency of each high-frequency word in each blog Inverse document frequency; based on the blog's non-dictionary word density, blog malicious tendencies, and the frequency of each high-frequency word in each blog-Inverse document frequency, and codingbased on the correlation or uncorrelation between blog and IOC to train the preset classification model to obtain a blog classifier; gettomg the non-dictionary word density of the blog to be categorized, blog malice, and the frequency of high-frequency words Reverse the document frequency and input to the trained blog classifier to obtain the classifier output value indicating that the blog to beclassified is related or not related to the IOC. Through the above method, IOC-related blogs and IOC-independent blogs in network security technology blogs can be accurately classified.

The invention relates to a dynamic classification analysis method and a system for enterprises. The product type is redefined through clustering, the candidate calculation dimension is reduced, and the calculation overhead is reduced; the actual operation state of the enterprise is ingeniously fused into the representation vector of the enterprise; according to the method, scattered report information without comparability is converted into enterprise vector data which can be directly calculated and a gap of cross-domain data types, so that mechanical rapid processing of the data is realized,and the data analysis efficiency is improved.

The invention discloses a video classification method and device, electronic equipment and a storage medium, and particularly relates to the technical field of artificial intelligence such as computer vision and deep learning. The specific implementation scheme is as follows: acquiring multiple groups of image frames from a to-be-processed video; performing feature extraction on each group of extracted image frames to determine a feature map corresponding to each group of image frames; determining a weight corresponding to each group of image frames; based on the weight corresponding to each group of image frames, fusing the feature maps to generate a total feature map corresponding to the video; and classifying the videos according to the total feature map to determine category labels corresponding to the videos. Therefore, when the video is classified, not only is the content contained in the image frame in the video considered, but also the effect of the image frame in the video is considered, so that the classification result is more accurate and more reliable.

The present invention discloses a polarization SAR image classification method based on non-stationary modeling of a polarization scattering mechanism, in order to solve the problems that the existingpolarization SAR image classification is affected by noise and has low accuracy for the mixed pixels with no obvious main scattering mechanism. The implementation steps are: initially classifying measured images; estimating the auxiliary random field according to the polarization scattering characteristics, and associating the polarization scattering characteristics with the non-stationarity; dividing the pixel point stationarity by using the auxiliary random field; calculating correlation functions for the stationary pixel points to obtain a unitary potential energy function, a data item, and a binary potential energy function; calculating the membership degree for non-stationary pixel points; constructing a posterior probability model of a fuzzy triple-recognition random field (FTDF) model by using the obtained functions, and performing classification by using the maximum posterior probability criterion; and if it is marked that the random field converges, outputting a result, and otherwise repeatedly constructing the classification model according the iterative rule until the termination iteration requirement is reached, and outputting a classification result. The method disclosed by the present invention has high detection precision and good anti-noise performance, and can be used for polarization SAR image classification.

The invention discloses a ship trajectory classification method based on feature selection and hyper-parameter optimization. For the problem of unbalanced original data in ship trajectory classification, firstly, data cleaning and preprocessing are performed on multiple pieces of trajectory data of an automatic ship identification system, then all trajectories are drawn by matplotlib, unavailable trajectories are deleted, then features are calculated to obtain additional features, dimensionality reduction processing is performed on speed, course and trajectory coordinates, backward feature selection is carried out on all features after dimension reduction processing through a random forest (RF), finally hyper-parameter optimization is carried out through the random forest, and classification prediction is carried out on the features through model training and performance evaluation without depending on an external information source. The ship trajectory classification method has high performance and is stable, and can be effectively applied to actual ship trajectory classification.

The invention discloses a public opinion information emotion classification method. The method comprises the following steps: collecting a certain amount of public opinion information data and carrying out emotion labeling; collecting the public opinion information data needing to be predicted; training a Google BERT model by using the marked training set; utilizing a Google BERT model to classify emotion types of the unlabeled public opinion information data; and visually displaying the public opinion information processing result through echarts, wherein the emotion classification accuracy of the method on a test set can reach 82%, and the processing result can be flexibly and visually displayed in a visual mode.

The invention provides an out-of-distribution image detection method based on attention enhancement and input disturbance, the method adopts an input disturbance technique, the influence on samples in distribution is greater than that of samples out of distribution, so that the confidence score of the samples in distribution is higher, and meanwhile, a temperature scaling technique is used, the prediction probability distribution of the samples in the distribution is more sharp, the prediction probability of the samples out of the distribution is smoother, and the confidence score difference between the samples inside and outside the distribution is further increased; compared with the mode that a generative model is directly used for carrying out a distributed external sample detection task, the method has the advantages that the method does not need to introduce additional hyper-parameters, the model is relatively simple, and the training time can be saved; compared with a generative adversarial method for performing an out-of-distribution sample detection task, the method has the advantages that the method is not excessively limited to training data, misjudgment is not easy to generate for edge samples, and a better detection effect can be obtained.

The invention relates to a nuclear magnetic resonance image classification diagnosis method for adult diffuse glioma. The method comprises the following specific steps: step 1, respectively acquiring T2-fluid attenuation inversion recovery T2-FLAIR and contrast enhancement T1 weighted imaging CE-T1WI magnetic resonance images of glioma labeled by a doctor according to histological and genomics information; step 2, carrying out data preprocessing operation on the T2-FLAIR image data and the CE-T1WI image data; step 3, each pair of T2-FLAIR and CE-T1WI images after preprocessing is sent to a decision tree diagnosis model, and an Attention-DenseNet model is used for learning and extracting the characteristics of the T2-FLAIR and CE-T1WI images; 4, obtaining a prediction result based on a plurality of independently trained Attention-DenseNet models, fusing the outputs of the plurality of models by adopting a voting mechanism or a weighted average strategy, and generating a final classification decision; and 5, sending the obtained final classification to a platform, and making final diagnosis based on a classification result.

The invention provides an individual emotion recognition method fusing expressions and postures, and mainly relates to classification of individual sequence emotions by using a multi-mode individual emotion recognition network. The method comprises the following steps: constructing a multi-modal individual emotion recognition network (including two channels for processing an expression sequence and a posture sequence), extracting expression features and posture features in a video sequence in parallel by using the network, and finally fusing the two features to obtain individual sequence emotion classification. According to the method, the self-learning ability of deep learning is fully exerted, the limitation of manual feature extraction is avoided, and the adaptive capacity of the method is higher. And the structural features of the multi-stream deep learning network are used for parallel training and prediction, and finally classification results of multiple sub-networks are fused, so that the accuracy and the working efficiency are improved.

The invention provides a squid identifying and sorting method which comprises the following steps of firstly, extracting a squid area from an image containing squids, then judging the orientation andintegrity of the squids through the area characteristics, extracting the necks of the complete squids through a morphological algorithm, calculating the body length, and dividing the squids into a plurality of grades according to the body length; for the incomplete squids, marking the squids directly as the lowest level; and according to the grading condition, controlling the external selection equipment to distribute the squids, thereby finishing classification. According to the method, the squids can be automatically identified and graded, the grabbing and distribution are controlled, the manual operation is not needed during the whole process, the intelligent degree is high, the manpower is greatly saved, the identification accuracy is high, and the classification result is reliable. The method can accurately and effectively sort the different types of squids and the aquatic products, such as cuttlefishes, octopuses, etc., which have the similar shapes with the squids.

The invention relates to the technical field of tobacco leaf raw material quality characteristic analysis, in particular to a tobacco leaf raw material quality characteristic analysis method based on data fusion, which comprises the following steps: obtaining related indication data of tobacco leaves from different types of tobacco leaf samples, and converting the related indication data of the tobacco leaves into a related indication data matrix of the tobacco leaves; measuring infrared spectrum data of different types of tobacco leaf samples, preprocessing the measured infrared spectrum data of the tobacco leaf samples, performing principal component analysis on the preprocessed infrared spectrum data of the tobacco leaf samples, determining a principal component number k, and obtaining score P data and load V data; combining the score P data matrix with each related indication data matrix of the tobacco leaves to obtain a combined tobacco leaf quality data matrix D, and normalizing the tobacco leaf quality data matrix D to obtain a combined tobacco leaf quality data matrix X; and carrying out clustering analysis processing on the fused tobacco leaf quality data matrix X, and analyzing quality characteristics of different types of samples.

The invention discloses a manhole cover disease detection method based on virtual sample driving, and the method comprises the steps: generating a large number of virtual samples simulating real manhole cover diseases through AIGC, i.e., a generative artificial intelligence technology, constructing a diseased manhole cover virtual data set, carrying out the training of a target detection model, and carrying out the discovery and recognition of the manhole cover diseases in a real image through employing the model; a virtual data set is constructed by using a virtual image generated by AIGC, so that the long-tail distribution phenomenon of well lid disease types is effectively inhibited, and well lid disease labels are smoothed; in combination with posterior knowledge, model optimization is guided, that is, after model training, the model is further utilized to carry out index analysis on a real image recognition result, and feature distribution in a targeted smooth virtual data set is carried out from three dimensions of a disease well lid scene easy to recognize by mistake, a disease form easy to recognize by mistake and a disease well lid type easy to recognize by mistake and omission; correcting and optimizing a model prediction result; through the method, the model can identify well lid diseases more accurately, and more reliable disease classification and statistical results are provided.

The embodiment of the invention provides an image processing method and device, equipment and a medium, and the method comprises the steps: carrying out feature extraction of a to-be-processed image,and obtaining a feature map of the to-be-processed image; determining an attribute prediction value of each feature point on the feature map, wherein the attribute prediction value of one feature point represents the probability that a pixel point corresponding to the feature point in the to-be-processed image belongs to a text region, the prediction distance between the pixel point and a plurality of position points of the text region to which the pixel point belongs, and the prediction category of the text region to which the pixel point belongs; and according to the attribute prediction value of each feature point on the feature map, marking the position and category of each text region on the to-be-processed image. By adopting the technical scheme of the invention, the accuracy of positioning and classifying the text area in the image can be improved.

The invention discloses an image classification method based on deep learning multi-scale feature fusion, which utilizes a deep learning multi-scale feature fusion method to improve the accuracy of image classification by combining self-calibration convolution, an attention module and a feature adjustment module. Firstly, feature extraction is carried out on an image through the self-calibration convolution module, then the attention module enhances the features, important areas in the image are highlighted, and finally the feature adjustment module adjusts the enhanced features to improve the expressive power of the features. Besides, through multi-time iteration enhancement and adjustment and adoption of various loss function optimization features, complex image data can be processed and classified more accurately, and the method is suitable for various application scenes requiring high classification precision. Compared with existing methods such as ResNet and Vision Transformer, the method has the advantages that the loss value is lower, the classification accuracy is higher, and excellent performance and wide application prospects are shown in the field of image classification.

The invention discloses a multi-source unstructured data cleaning method for discrete intelligent manufacturing application, and the method comprises the steps: carrying out the characterization analysis of multi-source unstructured data in a discrete intelligent manufacturing application environment, and classifying the cleaning types; and carrying out data cleaning on the to-be-cleaned multi-source unstructured data according to the data cleaning strategy corresponding to the cleaning type. The problem of unified description of the multi-source unstructured data and the problem of complexity of data classification processing are solved; cleaning type classification processing of multi-source unstructured data applied to discrete intelligent manufacturing by means of a computer becomes possible, computer processing is short in time consumption and has certain high efficiency, the cleaning types of the multi-source unstructured data are reflected by adopting the cloud model, the problems of unclear expression of fuzzy cleaning types and the like are avoided, therefore, the classification result of the cleaning types is more reliable, and a new technical solution is provided for multi-source unstructured data cleaning of discrete intelligent manufacturing application.

The invention discloses a ship trajectory clustering method based on a curve length distance, and relates to the technical field of ship navigation trajectory analysis. The method mainly comprises the steps: obtaining a weighted length after the normalization of a curve length between a tail end trajectory point in a current trajectory segment and a head end trajectory point of a corresponding ship trajectory; according to the proportion of the weighted length of the current tail end trajectory point in the corresponding ship trajectory, obtaining trajectory points at the same weighted length proportion in other ship trajectories through a linear interpolation method to serve as paired trajectory points; calculating comprehensive similarity measurement according to the navigation data of the current tail end track point and the corresponding paired track point; obtaining a clustering result through a DBSCAN clustering method; and extracting a ship typical channel. According to the method, physical information contained in the channel is mined through multi-dimensional data in an existing ship AIS, the track clusters are formed through clustering of ship tracks to represent ship behavior characteristics, and therefore the track clustering method small in difference interference and low in complexity can be achieved.