305 results about "Neural network classification" patented technology

The invention discloses a CT image pulmonary nodule detection system based on 3D full-connection convolution neural network. The detection system comprises the following five steps: constructing a training set data; performing 3D convolution neural network classification network training; performing 3D convolution neural network segmentation network training; carrying out false-positive suppression in which the trained segmentation network and the false-positive are utilized to inhibit the network; and detecting the pulmonary nodule. The technical schemes of the invention can realize the full automatic detection without any human intervention. At the same time, the recall rate of pulmonary nodule detection can be increased effectively; the false-positive focus of infection is reduced considerably; and a pixel level positioning quantitative and qualitative result for the focus-of-infection area of the pulmonary nodule can be obtained.

A wavelet-based band difference-sum ratio method reduces the computation cost of classification and feature extraction (identification) tasks. A Generalized Difference Feature Index (GDFI), computed using wavelets such as Daubechies wavelets, is employed in a method to automatically generate a large sequence of generalized band ratio images. In select embodiments of the present invention, judicious data mining of the large set of GDFI bands produces a small subset of GDFI bands suitable to identify specific Terrain Category / Classification (TERCAT) features. Other wavelets, such as Vaidyanathan, Coiflet, Beylkin, and Symmlet and the like may be employed in select embodiments. The classification and feature extraction (identification) performance of the band ratio method of the present invention is comparable to that obtained with the same or similar data sets using much more sophisticated methods such as discriminants, neural net classification, endmember Gibbs-based partitioning, and genetic algorithms.

The invention discloses a knowledge graph representation learning method. The method comprises the following steps: defining the correlation between entity vectors and relation vectors in a relation triple (head, relation, tail) by utilizing a translation-based model between the entity vector and the relation vector; defining the correlation between entity vectors and feature vectors in a feature triple (entity, attribute, value) by utilizing a neural network classification model; and correlating the entity vectors, the relation vectors and the feature vectors with one another through an evaluation function, and minimizing the evaluation function to learn the entity vectors, the relation vectors and the feature vectors so as to achieve the optimization aim. By adopting the method disclosed in the invention, the relation among the entity, the relation and the feature can be accurately represented.

The invention proposes a garment classification and collocation recommending method and a garment classification and collocation recommending system based on a deep convolution neural network. The method comprises the following steps: adding batch-normalized and improved inception structures, adding a redundant classifier to improve an original GoogleNet convolution neural network, and extracting the features of garment images to get the classification result of the garment images; performing image augmentation on a collocation library training set in multiple ways, distorting and turning the garment images and transforming the color space of the garment images, and training an improved GoogleNet convolution neural network classification model; and looking for similar items and collocations thereof in a suit library, generating identity information for each garment image, comparing the identity information to get similar images, and recommending garment collocations according to the gender, style and function information corresponding to the garment images. Corresponding garment collocation advices can be given to consumers according to input garment images. The method and the system have the advantages of high speed and high precision.

A method for identifying malicious encrypted network traffic associated with a malware software component communicating via a network, the method including: defining, for the malware, a portion of network traffic including a plurality of contiguous bytes occurring at a predefined offset in a network communication of the malware; extracting the defined portion of network traffic for each of a plurality of disparate network connections for the malware; evaluating a metric for each byte in each extracted portion; representing each extracted portion in a matrix data structure as an image of pixels wherein each pixel corresponds to a byte of the extracted portion; training a neural network based on the images for the extracted portions such that subsequent network traffic can be classified by the neural network to identify malicious network traffic associated with the malware based on an image generated to represent the defined portion of the subsequent network traffic.

The invention discloses a deep convolutional neutral network and superpixel-based image semantic segmentation method, which overcomes the problem that the precision of an existing semantic segmentation method still needs to be improved in the prior art. The method comprises the following steps of 1, training a deep convolutional neutral network classification model from images to category labels on an image classification data set; 2, adding a deconvolutional layer to the deep convolutional neutral network classification model, performing fine adjustment training on an image semantic segmentation data set, and realizing mapping from images to image semantic segmentation results; 3, inputting test images to a deep convolutional neutral network semantic segmentation model to obtain semantic labels of pixels, and inputting the test images to a superpixel segmentation algorithm to obtain a plurality of superpixel regions; and 4, fusing superpixels and the semantic labels to obtain a final improved semantic segmentation result. The method improves the precision of the existing semantic segmentation method and is of important significance in image identification and application.

The present invention discloses post-wavelet analysis treating method and device for electric power transient signal. The treating method for electric power transient signal after wavelet analysis and before feeding to the electric power monitoring center includes the following treatment on wavelet coefficient: the extraction of module maximum and the detection of irregularity; the statistics and cluster analysis of wavelet coefficient; neural network classification; energy analysis; and wavelet entropy calculation. The present invention can extract the characteristic of electric power transient signal for the application in traveling wave ranging, fault recognition, electric energy quality analysis and equipment fault diagnosis in the transmission line of power system.

The invention discloses a built-in drug target interaction prediction method based on a heterogeneous network. The built-in drug target interaction prediction method includes the steps: based on the assumption that a chemically similar drug can often interact with a similar target, combining a drug-drug similarity network, a target-target similarity network, and a drug-target interaction network into a drug-target heterogeneous network; using a starting-node-based migrating sequence, constructing a neural network classification model, taking the migrating sequence as input of the neural network classification model, training the classification model and learning to obtain vector representation of all nodes; and for prediction of drug-target interaction, giving a pair of drug-target pairs,extracting vector representation of the corresponding drug and target from the node vectors obtained from learning, performing Hadamard product operation on the two vectors, and taking the obtained result as the input of a random forest classifier to obtain the final prediction result. According to the experimental verification, the built-in drug target interaction prediction method based on a heterogeneous network has preferable prediction effect and applicability.

Genetically adaptive neural network systems and methods provide environmentally adaptable classification algorithms for use, among other things, in multi-static active sonar classification. Classification training occurs in-situ with data acquired at the onset of data collection to improve the classification of sonar energy detections in difficult littoral environments. Accordingly, in-situ training sets are developed while the training process is supervised and refined. Candidate weights vectors evolve through genetic-based search procedures, and the fitness of candidate weight vectors is evaluated. Feature vectors of interest may be classified using multiple neural networks and statistical averaging techniques to provide accurate and reliable signal classification.

The embodiments of the present invention provide a target detection method and device. The method includes the following steps that: an image to be detected is acquired; a plurality of candidate areasof the image to be detected are classified according to a cascade neural network, at least one level of neural network of neural networks starting from the second-level neural network includes a plurality of parallel sub neural networks of the corresponding level, wherein the sub-neural networks classify classification results of a previous level of neural network; and a target area is determinedaccording to the final classification results of the plurality of candidate areas. According to the method and device provided by the embodiments of the present invention, at least one level of neural network of the neural networks starting from the second-level neural network includes the plurality of parallel sub neural networks at the corresponding level, so that the candidate areas can be classified more comprehensively and accurately, and therefore, classification accuracy can be improved, and the target area can be accurately determined; and the reduction of the neural networks can be benefitted, and storage space occupied by a classification model composed of various levels of neural networks can be decreased. The method and device can be applied to devices with low hardware configurations or low computing performance.

The invention relates to the technical field of hyperspectral image classification, and discloses a hyperspectral image deep learning classification method, device and equipment and a storage medium,which are used for improving the accuracy and efficiency of hyperspectral image classification. The method comprises the following steps: acquiring a to-be-classified hyperspectral image; carrying outrandom clipping on a to-be-classified hyperspectral image according to a preset window size and the marked sample set to obtain a to-be-trained sample set; expanding the data set through image transformation to obtain a corresponding deep learning sample set; extracting spatial spectrum features by adopting a convolutional recurrent neural network and a three-dimensional convolutional neural network; and classifying the hyperspectral images through a preset neural network classification model obtained through training to obtain a corresponding image classification result. By constructing thedeep neural network model, the deep abstract features of the hyperspectral image can be automatically extracted, the workload of manual feature extraction and optimization is effectively reduced, andthe end-to-end automatic identification and classification of the hyperspectral image are realized.

The invention relates to a device and a method for detecting sputum smear tubercle bacillus quickly. The device comprises a support, a charge coupled device (CCD) image sensor, a microscopic imaging system, a computer system, a three-dimensional accurate control console, an X-axis motor, a Y-axis motor, a Z-axis motor, wherein a high-accuracy grating ruler with a grating sensor is arranged in the Z-axis direction of the support additionally; a light-emitting diode (LED) light source plate of which the light intensity is controllable is positioned on the lower side of the microscopic imaging system and is used as an object stage; and the computer system comprises three software functional modules, namely a control module, an image target fusion module and a neural network classification and counting module. The method comprises a visual sense area interpretation algorithm and an optimum detection path scanning algorithm which are used for image detection and a method for fusing targets twice in image segmentation. By the device and the method, the detection accuracy and efficiency of machine vision can be improved, and the scientific reference is provided for the clinical diagnosis and prevention and treatment of tuberculosis.

The invention discloses a high reflection surface defect detection method based on image processing and neural network classification. The method includes: performing basic processing including background deduction and denoising on an originally acquired image, preliminarily determining positions of parts possibly having defects through feature extraction and obtaining a series of area images of the corresponding positions, inputting an image sequence to a neural network classifier regarding a local feature area block as training data, determining whether the defects are real defects, and regarding an output result of the classifier as a final determination result. According to the high reflection surface defect detection method, defect feature search and extraction are performed by employing a front-end two-dimensional digital image processing module, feature filtering and enhancement are performed with the combination of a rear-end neural network classifier, the accuracy of the search result is enhanced, surface defects of a detected member are fully extracted, the probability of false detection and omitted detection in the conventional image processing detection method is reduced, and the operation efficiency and the versatility are simultaneously considered.

The dynamic characteristic analysis method of real-time tendency of heart state includes: obtaining electorcardiac waveform with multipath electorcardiac amplifier and 12-bit A / D converter; forming time sequence related heart rate variation scatter diagram by means of space-time correlation technology and scatter diagram technology; extracting time sequence related characteristic parameters, short time-real time characteristic conversion illustration parameter and quantized space-time parameter indexes of characteristic illustration via automatic and manual interaction; classifying the illustration and quantized space-time parameters in artificial neural network; and describing dynamic characteristics and relevant invariance characteristics of heart rate variation in a nine-dimensional and a five-dimensional space with two independent curve surfaces and their boundary to represent the heart function.

The invention discloses a device and method for collecting and segmenting of hyperspectral images of unstainedpathological sections. A section sample platform is supported in the middle of a bracket, the hyperspectral images of the unstainedpathological sectionsare automatically collected and processed through a computer, and lesion area segmentation resultsare obtained. Based on spectral difference caused by pathological changes of tissues, a personal computeris adopted to synchronously control relevant modules, spectral sequence images of the unstainedhistopathological sectionsare collected and preprocessed, and corresponding three-dimensionalhyperspectral dataare produced after overlaying;based the data and the combination with a currentlypopularneural network classificationthought, a spectral classification algorithm is developed to perform identity partitioning of a lesion area, thereby acceleratingthe identification rate and efficiency of the histopathological section, avoidingmanual errors which maybe brought during staining, and reducingthe time required by section making;automatic judgment is performed by utilizing a machine algorithm, therebyreducingsubjectivity caused by manual judgment. Therefore,a good auxiliary function is provided for a pathology doctor to detect the pathological sections.

The invention relates to a method and a system for identifying the white-leg shrimp disease on the basis of machine vision. The method comprises the following steps of: S1, judging whether an image is an image of a target to be subjected to disease identification, entering the step S2 if judging that the image is the image of the target to be subjected to disease identification, and stopping a program if judging that the image is not the image of the target to be subjected to disease identification; S2, extracting a color feature parameter of the image; S3, carrying out binary segmentation processing on the image; S4, extracting an area feature of the image which is subjected to binary segmentation processing, and calculating the number of pixel points in a target region; S5, carrying out edge detection processing on the image which is subjected to binary segmentation processing to obtain an edge image of the target region, then extracting a perimeter feature of the edge image and obtaining the number of pixels in a target edge region; S6, obtaining a circularity feature parameter by utilizing a ratio of the perimeter to the area of the target region; and S7, obtaining a disease identification result by training the color feature parameter and the circularity feature parameter which are used as training parameters and categorical data sources of a neural network classification algorithm and then classifying the color feature parameter and the circularity feature parameter.

The invention discloses an intelligent graphic retrieval method, which is characterized in that: extracting features of graphics to generate feature set by the method of Fourier change, training RBF neural network classification model by taking one part of the feature set as training set, indexing the graphics using the classification result given by the classification model; client of a retrieval system extracts the features of retrieval graphics, gives a category by the trained classification model and computes the similarity distance between the retrieval graphics and each graphics in the feature set of the same category; sorting the similarity distance, returning to the graphics according to the number made by the system and further revising RBF neural network classification model by relevant feedback methods. The invention improves the intelligence of search process, effectively determines the RBF neural network classification model by improved algorithm of subtractive clustering, greatly improves retrieval precision, speeds up retrieval speed and upgrades retrieval performance.

The invention provides a method for diagnosing faults of urban gas pipelines based on a deep learning neural network. According to the method, a deep learning neural network classification model formed by a sparse automatic encoder SAE and SOFTMAX is established and is combined with a pipeline fault type so as to construct a fault diagnosis model, and then the fault classification of the gas pipelines is realized; and by virtue of unsupervised automatic learning characteristic parameters in deep learning and a supervised fine adjustment network, the problems that the characteristic parametersof fault diagnosis of the pipeline are based on experiences and the diagnosis accuracy rate is low are effectively solved, the fault diagnosis of the urban gas pipelines in the working process is relatively rapid, and the stability, precision and reliability of the fault diagnosis are improved.

The invention provides a gas pipeline leakage identification method based on a convolution neural network. The method comprise the following steps that after a leakage sound signal and a background sound signal of a typical leakage type are collected, framing processing and short-time Fourier transform are carried out to obtain a time-frequency diagram representing the original leakage sound signal; then a convolution neural network classification model aiming at the leakage sound signal is built, a traditional square convolution kernel is changed into a specific strip-shaped rectangular convolution kernel, so that the line spectrum characteristics in the time-frequency diagram are better extracted; and the time-frequency diagram of the leakage sound and the time-frequency diagram of the background sound are mixed and sent to the built convolution neural network for training, K-fold cross validation is adopted for training, and a network model superparameter is optimized, so that the optimal model superparameter is selected and the robustness and universality of the model are enhanced. Compared with the pipeline leakage identification method in the prior art, the method has the advantages that the identification rate is further improved, and the problem of feature screening which is most difficult to process in the prior art can be effectively solved.

The invention discloses a face direction change recognition method based on a neural network and sensitivity parameters. The face direction change recognition method comprises the steps of: carrying out first static face orientation recognition on acquired single-frame color images one by one, including preprocessing the single-frame color images and extracting facial feature vectors, and judging whether the face orientation of each single-frame color image is frontal, leftward or rightward according to positions of eyes and / or nose of the facial feature vectors; carrying out the first static face orientation recognition on all the acquired single-frame color images within given acquisition time, so as to obtain a first face orientation result set with results arranged in turn according to time sequence and a plurality of facial feature vectors; and adopting neural networks classification for carrying out process analysis on the plurality of facial feature vectors and the first face orientation result set, recognizing instruction intention, and acquiring a first instruction result given in the face direction change process. The face direction change recognition method can achieve accurate face direction change recognition under the condition of strong backlight of the acquired images.

The invention discloses a fingerprint detection classification method based on the space transformation convolutional neural network. The fingerprint detection classification method comprises a fingerprint image extraction region of interest preprocessing, image high-frequency region extraction, image space transformation processing and convolution neural network classification training and testing. The fingerprint image extraction region of interest preprocessing removes a blank region through extracting a fingerprint part in an image; the high-frequency region extraction means the high-frequency characteristic of the image is extracted through a gaussian high-pass filter; as for the image space transformation processing, the space transformation neural network is used for carrying out translation, cutting and rotating operation on the input image, so that expansion of image data is achieved; the convolution neural network adopts multi-layer convolution pooling, convolution kernels with different sizes are used for extracting image features, and a good classification detection effect is obtained on the test set. The invention provides a fingerprint detection method which is low incost, high in detection precision and short in time consumption.

The invention discloses a health condition monitoring method of complex equipment based on an attention mechanism and a neural network. The main steps comprise acquiring tge multi-sensor data of complex equipment; obtaining the effective measurement data by feature selection; obtaining a plurality of slice samples by preprocessing; establishing the neural network classification model which combines the attention mechanism and the depth neural network; inputting the slice samples and their corresponding labels into the neural network classification model to train the neural network classification model offline; inputting the slice samples of multi-sensor data to be predicted into the trained neural network classification model to obtain the health status of complex equipment. The method considers the data characteristics of the multi-sensor signal, fully excavates the local characteristics and the time sequence information in the data, has high prediction accuracy and wide applicability, and can be widely applied to various complex equipment.

The invention discloses an intelligent disease inquiry method, device and equipment and a storage medium, and the method comprises the steps: obtaining a natural inquiry statement input by a user, andextracting symptom keywords in the natural inquiry statement; identifying an inquiry category in the natural inquiry statement according to a preset neural network classification model; when the inquiry category is triage inquiry, inputting the symptom keyword into a preset triage model, and obtaining a triage department which is output by the triage model and corresponds to the symptom keyword;and when the inquiry category is symptom inquiry, inputting the disease keyword into a preset symptom question and answer model, and obtaining a disease entity which is output by the disease questionand answer model and corresponds to the disease keyword. According to the method, the triage precision can be improved, the diagnosis time wasted by a doctor due to the fact that the doctor hangs thedepartment number by mistake is saved, the hospitalizing efficiency is optimized, meanwhile, the patient can conduct disease inquiry to have more understanding on own diseases or symptoms, and the user experience is greatly improved.

The invention discloses a precise reservoir prediction method based on waveform classification and retrieval under forward constraints. The method comprises the steps of selecting an effective range and converting large-region multi-phase spread into micro-region single phase spread; dividing a reservoir in the single phase spread into different types, and summarizing sedimentary characteristics of different types of reservoirs and post-stack seismic reflection characteristics; simulating a typical waveform of each type of reservoirs via wave equation forward modeling; performing frequency expanding treatment on the single phase spread seismic data by using well control mixed phase wavelet deconvolution; performing primary waveform classification on a single phase spread range by using a non-supervision neural network classification method; and reconstructing a waveform model trace, re-classifying the waveforms, and implementing waveform retrieval. According to the method provided by the invention, a corresponding relation between the reservoir types and the waveforms is built by wave equation forward modeling, and the waveforms are primarily classified and then retrieved to achieve precise prediction of the different types of reservoirs in the phase spread, so that the method is the effective and rapid precise prediction technology for recognizing the different types of reservoirs in the same phase spread in a small range.

The embodiment of the invention provides a data processing method. The data processing method comprises the following steps: acquiring a plurality of sample images; Adding a label to the plurality ofsample images, adding a positive sample label to the sample image including a predetermined feature, and adding a negative sample label to the sample image not including the predetermined feature; Establishing a neural network classification model based on an attention mechanism; And training the neural network classification model by using the sample image added with the label to obtain an optimal classification model. According to the scheme, an attention mechanism is introduced into a neural network classification model as an initial training model; A neural network classification model with an attention mechanism introduced in the training process can extract discriminative features which are more beneficial to distinguishing whether the image is a positive sample or a negative sample,and then an optimal classification model which can more sensitively and accurately judge whether the image contains predetermined features or not is obtained. The embodiment of the invention furtherprovides a data processing device, a medium and computing equipment.

The invention relates to an object-neural-network-oriented high-resolution remote-sensing image classifying method, aiming at solving the problems that the conventional remote-sensing image classifying method is low in classification precision and cannot effectively utilize information of all wave bands of a remote sensor. The method comprises the following steps that: an image of the ground is shot by a high-spatial-resolution sensor and is transmitted to a computer; the computer carries out primary image element division on the input image by a region growing algorithm; the primarily-divided image is subjected to multi-size division according to continuously-set neterogeny degree thresholds and shape features and spectral signatures of the image, thus forming divided images with different sizes; and the obtained divided images with different sizes are used for establishing a BP (Back Propagation) neural network, setting training parameters and establishing training samples to classify the image which is subjected to the multi-size division, thus obtaining a high-resolution image. The method is applicable to the field of obtaining of images with high spatial resolutions.

A method for identifying malicious encrypted network traffic associated with a malware software component communicating via a network, the method comprising: defining, for the malware, a portion of network traffic including a plurality of contiguous bytes occurring at a predefined offset in a network communication of the malware; extracting the defined portion of network traffic for each of a plurality of disparate network connections for the malware; evaluating a metric for each byte in each extracted portion; representing each extracted portion in a matrix data structure as an image of pixels wherein each pixel corresponds to a byte of the extracted portion; training a neural network based on the images for the extracted portions such that subsequent network traffic can be classified bythe neural network to identify malicious network traffic associated with the malware based on an image generated to represent the defined portion of the subsequent network traffic.

The invention discloses a bearing state noise diagnosis algorithm facing network variable screening and characteristic entropy fusion. A sensor collects bearing operation noise signals. The noise signals are segmented according to a time sequence and form a sample set. A time-frequency domain characteristic of a sample is extracted so as to acquire a time-frequency-domain one-dimensional characteristic row vector. An average influence value algorithm is adopted to realize first characteristic variable screening so as to acquire a sensitive characteristic set, and through calculating a characteristic entropy of the sensitive characteristic set, characteristic secondary screening and dimensionality reduction are performed on an average influence value similarity characteristic so as to acquire a final characteristic set. A PSO or GA optimization support vector machine is used to carry out training and establish a fault diagnosis model so as to determine a bearing fault type and output aresult. In the invention, complementarity of a characteristic average influence value and the characteristic entropy based on a network in characteristic selection and characteristic classification isused; and a disadvantage that the characteristic selection and a neural network classification algorithm are mutually isolated in bearing noise diagnosis is overcome so that a time-frequency domain characteristic index well reflects a bearing operation state and a classification network characteristic.

The invention provides a method for classifying types of mixed seabed sediment based on a multi-beam sonar technology, and belongs to the fields of marine acoustic remote sensing detection and recognition. The method comprises the following steps: according to backscatter strength data obtained by a multi-beam sonar system, improving the existing multi-beam backscatter strength data correction model by analyzing influence on backscatter strength due to factors such as submarine topographical features, reflected signals at a central wave bundle area and the like; on the basis of the improved model, systematically seeking relationship between the submarine backscatter strength and sediment types and characteristics in details in combination with real submarine sediment sample data obtained through submarine geological sampling; and recognizing the types of the mixed seabed sediment rapidly, accurately and automatically by a neural network classification method. The method provided by the invention has the advantages of strong practicability and strong generality, and is mainly used for classifying and recognizing the types of the mixed seabed sediment.

An early recognition method of thermal power plant steam feed pump fault features depend on the supervision information system (SIS) of a thermal power plant configured by a lot of thermal power plants, and through joining and processing of historical data of related measurement points of a steam feed pump set and the typical fault maintenance record of the steam feed pump set in an SIS database, the neural network classification method in the data mining algorithm is employed to realize the classification of the data features prior to the generation of the typical faults of the steam feed pump set so as to realize that a classification model can perform identification of fault features and perform timely alarm at the early phase when the fault features are represented in the data to select time to perform maintenance and avoid generation of accidents or unsafe events caused by device faults. The early recognition method of the thermal power plant steam feed pump fault features depends on the SIS, is convenient to use and high in practicality, finds new values for massive historical data stored by the SIS of the thermal power plant and explores new direction for the state maintenance of the steam feed pump set even the thermal power plant or more devices or systems.