477results about How to "The classification result is accurate" patented technology

The invention discloses a Chinese network review emotion classification method based on an integrated study frame. According to the method, a part-of-speech combination mode, an order-preserving sub-matrix mode and a frequent word sequence mode are adopted as input characteristics, in the level of characteristics, factors of the influence of Chinese word order information, interval phrase characteristics and the sentence length are considered, and the characteristic vector sparsity problem is solved through semantic similarities; the problem that many review text characteristics exist is solved, the inter-base-classifier independence is guaranteed, and the classification performance of base classifiers is improved as much as possible; a base classifier algorithm constructed based on product attributes is adopted to comprehensively review emotion information of each attribute in a text, and then the sentence-level emotional tendency of reviews is judged, so that a final classification result is more accurate. The Chinese network review emotion classification method based on the integrated study frame is applicable to e-commerce network review emotion classification in various fields, can make a potential consumer know evaluation information of a commodity before purchase and can also make a merchant better sufficiently know the consumer's opinion, and therefore the service quality is improved.

A method for classifying tissue in a magnetic resonance image. and particularly for measuring a volume of pathological tissue such as white tissue hyperintensity (leukoaraiosis) in the brain based on the segmentation of the intensity histogram of fluid attenuated inversion recovery (FLAIR) images is described. A magnetic resonance image of the brain of a subject is acquired, and a pixel intensity histogram is constructed from the image. A statistically-based regression analysis is applied to the histogram to determine upper and lower threshold values, which define different types of brain tissue, particularly normal brain, cerebral spinal fluid (CSF), or lesion.

A visual commerce engine can provide information related to an object based on an image of the object. The visual commerce engine receives from a user device an image of an object and a location within the image associated with the object, and analyzes the image to detect potential objects depicted in the image. From this set a detected object can be selected based on the received location's proximity to any of the detected potential objects. A description of the detected object can then be determined and compared with a library of objects to identify similar, identical, or related objects.

The invention discloses a deep learning-based question classification model training method and apparatus, and a question classification method and apparatus. The question classification model training method comprises the steps of extracting feature information samples in question text samples, and generating corresponding first eigenvector samples; performing spatial transformation on the first eigenvector samples to obtain second eigenvector samples; inputting the second eigenvector samples to a plurality of convolutional layers and a plurality of pooling layers in a multilayer convolutional neural network, and by superposing convolution operation and pooling operation, obtaining first fusion eigenvector samples; inputting the first fusion eigenvector samples to a full connection layer in the multilayer convolutional neural network to obtain global eigenvector samples; and training a Softmax classifier according to the global eigenvector samples to obtain a question classification model. The method can avoid a large amount of overheads of manual design of features; and through the question classification model, a more accurate classification result can be obtained, so that locating of standard question and answer is improved.

The invention discloses a vehicle recognition and tracking method based on convolutional neural networks. Through the method, the problem that it is difficult to guarantee instantaneity under a high-precision condition in the prior art is solved, and the defects of inaccurate classification results, long tracking and recognition time and the like are overcome. The method comprises the implementation steps that a quick region convolutional neural network is constructed and trained; an initial frame of a monitoring video is processed and recognized; a tracking convolutional neural network is trained off line; an optimal candidate box is extracted and selected; a sample queue is generated; online iterative training is performed; and a target image is acquired, and instant vehicle recognitionand tracking are realized. According to the method, a Faster-rcnn and the tracking convolutional neural network are combined, and high-level features with good robustness and high representativeness of vehicles are extracted by use of the convolutional neural networks; through network fusion and an online-offline training alternating mode, time needed for tracking and recognition is shortened on the basis of guaranteeing high precision; the recognition result is accurate, and tracking time is shorter; and the method can be used for cooperating with an ordinary camera to complete instant recognition and tracking of the vehicles.

Embodiments of the present invention provide a signal classification method and device, and encoding and decoding methods and devices. The encoding method includes: dividing a current frame into a low-frequency band signal and a high-frequency band signal; attenuating the high-frequency band signal or a to-be-encoded characteristic parameter of the high-frequency band signal according to an energy attenuation value of the low-frequency band signal, where the energy attenuation value indicates energy attenuation of the low-frequency band signal caused by encoding of the low-frequency band signal; and encoding the attenuated high-frequency band signal or the attenuated to-be-encoded characteristic parameter of the high-frequency band signal. The technical solutions according to the embodiments of the present invention can improve the effect of combining the low-frequency band signal and the high-frequency band signal at the decoder.

The invention discloses a photo classifying method and a photo classifying device. The photo classifying method comprises the following steps: judging whether a to-be-classified photo has a face area by adopting a face detection algorithm; if the to-be-classified photo has a face area, performing feature extraction on all to-be-classified photos with a face area by use of a lightweight deep model network, and obtaining the face recognition features of the to-be-classified photos and at least one face attribute information; clustering the to-be-classified photos based on at least one face attribute information to obtain multiple clusters; and clustering the to-be-classified photos of each cluster based on the face recognition features to obtain multiple classes of photo albums, wherein each class of photo album stores the to-be-classified photos of the same person. The photo classifying method can realize offline automatic classification of the to-be-classified photos, the classification result is accurate, and the classification speed is high.

The invention discloses a local spline embedding-based orthogonal semi-monitoring subspace image classification method. The method comprises the following steps of: 1) selecting n samples serving as training sets and the balance serving as testing sets from image data sets, wherein the training sets comprise marked data and unmarked data; 2) building an extra-class divergence matrix and an intra-class divergence matrix by using the marked data; (3) training data characteristic space distribution by using a whole and building a Laplacian matrix in a local spline embedding mode; 4) according to a local spline, embedding an orthogonal semi-monitoring subspace model, and searching a projection matrix to perform dimensionality reduction on the original high dimension characteristic; 5) building a classifier for the training samples after the dimensionality reduction by using a support vector machine; and 6) performing the dimensionality reduction on the testing sets by using the projection matrix and classifying the testing sets after the dimensionality reduction by using the classifier. In the method, the information, such as image sample marking, characteristic space distribution and the like, is fully utilized; potential semantic relevance among image data can be found out; and image semantics can be analyzed and expressed better.

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 discloses a method of image classification based on a visual attention model in the technical field of image classification. The method of the invention comprises the following steps: step 1: certain quantity of images are randomly selected from a selected image database to be used as training samples; step 2: a characteristic vector of each image based on the visual attention model and the overall scarcity is captured; step 3: a characteristic vector of each image based on the visual attention model and the overall scarcity is calculated to the image to be classified; and step 4: the characteristic vectors captured in step 2 and step 3 are transmitted to a classifier to be classified; finally, the result of the classification of the image to be classified is obtained. The method can classify the image by capturing visual characteristic characters at a high level from the characters at a bottom level so as to ensure more precise result of the classification.

The invention discloses a method for classifying polarimetric SAR (synthetic aperture radar) images on the basis of Cloude decomposition and K-wishart distribution and mainly solves the problem that the prior art is poor in classification effect and high in computation complexity. The method includes the steps of firstly, reading a polarimetric SAR image to be classified, subjecting each pixel of the image to Cloude decomposition to obtain an entropy H and a scattering angle alpha; secondly, initially partitioning the polarimetric SAR image according to the values of the entropy H and the scattering angle into eight classes; thirdly, subjecting results of eight classes of the whole polarimetric SAR image to iteration to obtain more accurate classification results. Compared with classic classification methods, the method is stricter in partitioning the polarimetric SAR image, better in classification effect, lower in computation complexity and applicable to terrain classification and target recognition of the polarimetric SAR images.

A fake fingerprint detection method based on markov random field (MRF) and support vector machine-k nearest neighbor (SVM-KNN) classification includes steps: (1) fingerprint image feature extracting: (1.1) first-order statistics (FOS), (1.2) a gray level co occurrence matrix (GLCM) and (1.3) an MRF; (2) SVM training: training the FOS and the GLCM feature vector and the MRF feature vector to obtain a model A and a model B; (3) SVM-KNN classification: (3.1) the SVM classification mechanism and (3.2) SVM-KNN classifier forming; and (4) decision fusion for true and false fingerprint detection. Presently related articles for fake fingerprint detection by aid of the GLCM and the MRF are not found, and the fake fingerprint detection method achieves the purpose of identifying true and false fingerprints by aid of physical structures of the two feature quantized fingerprint images. Experiment results prove that the false accept rate and the false reject rate of the algorithm are respectively 1.84% and 1.79%, and therefore the fake fingerprint detection method is high in accuracy and good in practicality.

The invention discloses a network user classification method and device. The network user classification method comprises the following steps of: generating behavior characteristic vectors of users based on the network behaviors of the users; clustering based on the behavior characteristic vectors of the users so as to construct a user-topic matrix; determining the webpage category to which each user belongs on a basis of the user-topic matrix, wherein the network behaviors of the users comprise network access behaviors and search behaviors, and each element value of the user-topic matrix respectively represents the correlation degree between each user and each topic. According to the method and device disclosed by the invention, on a basis of various types of data of the user search behaviors and access behaviors, overall user behaviors are covered within a smaller calculation cost by the adoption of a user-topic model, so that the user information is sufficient to use, and a more accurate user classification result is faster to obtain.

A text classification apparatus directed to a plurality of languages, includes a unit for extracting information for converting a word from non-classified (unlabeled) texts, in a plurality of languages, into a word sense, a unit for learning a classification knowledge at a word sense level after converting a word extracted from a labeled text into a word sense, a unit for learning a classification knowledge at a word level from the labeled text, a unit for learning the classification knowledge at the word level from the classification knowledge at the word sense level and information on a relation between words extracted from the unlabeled text, and a unit for combining the respective classification knowledges to assign a category.

The invention discloses a method and a system for identifying and classifying paper money. By utilizing the method and the system, a sample signal degradation model meeting the application requirements is established through using a statistical method by utilizing a large number of existing reliable samples easy to obtain, states of the paper money with brand new color, eight to nine into new color, seven to eight into new color, below seven into new color, varying degrees of soiling, varying degrees of mutilation, varying degrees of cracks and fold existing in regions, then the classifier study is carried out, a to-be-identified sample is classified and identified, and therefore, a classifying result is accurately obtained, and the cost and the efficiency for developing a paper money identifying product are also reduced.

The invention discloses a polarization synthetic aperture radar (SAR) image classification method based on spectral clustering. The polarization SAR image classification method mainly solves the problem that an existing non-supervision polarization SAR classification method is low in accuracy. The polarization SAR image classification method comprises the steps of extracting scattering entropy H of representation polarization SAR target characteristics to serve as an input characteristic space of a Mean Shift algorithm combining with space coordination information; diving in the characteristic space with the Mean Shift algorithm to obtain M areas; choosing representation points of all areas on the M areas to serve as spectral clustering input to spectrally divide all areas, and further finishing spectral clustering on all pixel points to obtain pre-classification results; and finally classifying the whole image obtained from the pre classification with a Wishart classifier capable of reflecting polarization SAR distribution characteristics in an iteration mode to obtain classification results. Tests show that the polarization SAR image classification method is good in image classification effect and can be applied to non-supervision classification on various polarization SAR images.

The invention discloses a correction-based K nearest neighbor text classification method, which comprises the following steps of: pre-processing texts, namely performing word segmentation on each document in a training text set, removing stop words, and performing projectized expression on the texts; selecting text characteristics, namely reducing dimensions of text vectors, and selecting document characteristics closely related with a document theme concept as few as possible; and finally, performing classification by using a correction-based K nearest neighbor text classification algorithm construction classifier to obtain a classification result. The method has accurate classification result.

The invention discloses a high-resolution remote sensing image crop classification method based on deep learning, and belongs to the technical field of satellite remote sensing image processing and application. The objective of the invention is to solve the problems of fuzzy ground object boundary, high algorithm complexity, neglected characteristic waveband information and the like in crop classification of an existing high-resolution remote sensing image, and the problem that crop classification cannot be carried out on the high-resolution remote sensing image by using an existing neural network method. According to the method, a 36-waveband combination mode is adopted, a VGG neural network result is designed, deep learning of the VGG neural network is utilized, and accurate classification of crop plots in a high-resolution remote sensing image with the resolution of 10 m is achieved through multiple iterations.

The invention provides a handwriting recognition method based on an encrypted neural network. The handwriting recognition method comprises the following steps: 1, preprocessing feature data and labeldata; 2, constructing a deep learning model, and training hyper-parameters; 3, performing homomorphic encryption on the preprocessed data; 4, improving matrix point multiplication and an activation function, and constructing an encrypted neural network; and 5, carrying out classification identification by utilizing an encryption neural network, and combining a homomorphic encryption algorithm witha neural network model to realize protection of data privacy and prevent leakage of data information, so that the method is mainly applied to the field of information security and artificial intelligence crossing.

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 a rock fragment size classification method based on multiple features and segmentation recorrection. The method comprises the following steps that (1) an original rock fragment size image is cut and blocked to generate subblock images of 64*43; (2) color feature extraction is carried out on the subblock images; (3) texture feature extraction is carried out on the subblock images; (4) feature selection is carried out on color features and texture features; (5) a classifier is designed through a support vector machine (SVM), and rough classification on the subblock images is completed; (6) watershed segmentation is carried out on the original rock fragment size image, and each obtained rock area is marked as blob; (7) on the basis of segmentation, a result obtained through rough classification is corrected and adjusted to improve classification performance and obtain a final classification result. According to the classification method, different classification algorithms are synthesized, efficiency and accuracy are high, cost is low, and the real-time performance requirement of a mining field is met.

Embodiments of the invention disclose a face detection method and apparatus, and terminal equipment, and are applied to the technical field of information processing. The face detection apparatus performs feature sampling according to feature information of an image in a face candidate frame of a to-be-processed picture and obtains a plurality of pieces of sampling feature information, then a detection score is obtained according to a preset calculation function and the plurality of pieces of sampling feature information, and finally, the image in the face candidate frame is classified according to the detection score, namely, a classification result whether the image in the face candidate frame is a face can be obtained. Therefore, the plurality of pieces of sampling feature information can be obtained through feature sampling, features of the image in the face candidate frame are selectively expressed, and the finally obtained classification result is accurate; and the detection score of the image in the face candidate frame is obtained through the plurality of pieces of sampling feature information, and the plurality of pieces of sampling feature information obtained through anyfeature sampling mode can have the same detection scores so that the precision of the finally obtained classification result is improved.

The invention provides an image classification method, a device and a system, which relate to the technical field of image processing. The method comprise inputting a target image to a pre-trained classification master network; among them, the classification master network being used to extract the key features of the target image, and fine-grained classification being carried out based on the keyfeatures. The key features are related to the object to be classified in the target image, and the key features are independent of the scene in the target image. Classification result of target imageis obtained by classification master network. The invention can be adapted to use a plurality of scenes and accurately classify images under different scenes, and has certain general applicability.

The invention discloses a mountain fire detection method based on intelligent image analysis. The method includes a training stage and a testing stage. An SLIC method is used for obtaining superpixels of an image. Through the method, pixels of the image can be fast clustered, fitting can be well performed on the edge of a target, after superpixel blocks are obtained, color and texture characteristics of the superpixel blocks are put forward, and mountain fire classification can be performed on the image according to the characteristics. Through a fire disaster detection method based on textures and colors, the position of a mountain fire can be fast and effectively detected, and a precise classification result can be obtained. The superpixels generated through SLLC can be well matched with the edge of the target, and the segmentation effect is ideal. The characteristics used for describing a superpixel area are simple, effective and low in complexity and can meet the instantaneity requirement of mountain fire detection.

The invention discloses a polarized SAR classification method based on deep learning of shallow-layer characteristics and T-matrix. The polarized SAR classification method based on deep learning of shallow-layer characteristics and T-matrix mainly solves the problem that the prior art is low in the classification correct rate for the same natural object having obvious difference on the scattering information and different natural objects having similar scattering information. The polarized SAR classification method based on deep learning of shallow-layer characteristics and T-matrix includes the following implementation steps: 1, filtering an original polarized SAR image; 2, extracting the polarized shallow-layer characteristics of the polarized SAR image after filtering; 3, fusing the shallow-layer characteristics with the polarized SAR data after filtering to construct training simples and a test samples; 4, learning the training samples through a convolution neural network; and 5, classifying the test samples by means of the convolution neural network which is obtained through learning, and obtaining the final polarized SAR natural object classification result. The polarized SAR classification method based on deep learning of shallow-layer characteristics and T-matrix is high in the classification correct rate of the polarized SAR target natural objects, has good experimental effect on classification of the natural object targets of a large area, and can be applied to target identification and natural object classification of a large scene.

The present invention relates to an airport target detection method for a high-resolution remote sensing image in a complex background. The method comprises: using a full-convergence network to perform saliency detection on a remote sensing image, and using the improved LSD algorithm to extract line features of the remote sensing image; extracting saliency features and linear features to respectively divide the target candidate region, and taking the region simultaneously satisfying the saliency features and the linear features as a target candidate region; and extracting depth features of thecorresponding image in a candidate region by using the convolution network, converting the two-dimensional matrix features with different sizes into one-dimensional features with the same length by using the ROI Pooling network, calculating the target probability and the positional offset of the candidate region through two independent fully connected networks, and detecting an airport target inthe remote sensing image. According to the method provided by the present invention, priori knowledge such as the airport saliency, the parallel straight runway and the like is used to generate a small number of candidate regions, the difficulty of the test can be significantly reduced, the extracted candidate regions are more accurate, and the accuracy of detection and the positional accuracy ofmarking the airport region are improved.

The invention discloses a method for classifying texture images on the basis of a local edge pattern. The method includes steps of inputting original texture images of the images to be classified; dividing the original texture images into n image blocks; respectively computing local edge pattern texture spectrum features of the original texture images and the n image blocks on the basis of m types of texture primitives with different sizes, and serially connecting the local edge pattern texture spectrum features with one another to obtain overall fusion local edge pattern texture spectrum features of the images to be classified; classifying the images to be classified into categories of training images with the minimum Canberra distances according to the overall fusion local edge pattern texture spectrum features of the images to be classified. Length and width pixels of the size of each texture primitive are even numbers, and the minimum texture primitive contains 2X2 pixels. The invention further discloses a system for classifying the texture images on the basis of the local edge pattern. The method and the system have the advantages that texture information acquired by the method and the system is rich and robust, and the texture image classification accuracy is high.

The invention provides a feature classification method for a liver ultrasonic image. The method comprises the steps of S1, for a to-be-processed ultrasonic image comprising a liver section / part, automatically extracting a liver capsule line from the ultrasonic image; S2, based on the extracted liver capsule line, selecting multiple sampling points, and generating a triple feature of each samplingpoint; S3, extracting each triple feature, and classifying the extracted triple features; and S4, according to a classification result of all the extracted triple features, determining the type of theultrasonic image. According to the method, classification results of all image blocks are combined to obtain an accurate classification result; the noise interference is reduced; automatic identification and classification are realized; and the labor cost is reduced.

The invention provides a video-based non-supervision abnormal event real-time detection method. Specific to the aims of reducing the quantity of inter-frame feature points and lowering the complexity in calculation of the feature points at the same time, the feature points are detected with an interval method, namely, a video is segmented, the feature points are detected on a first frame, and only tracking is required subsequently. The calculation amount of a tracking method is small relatively, so that the calculation complexity is lowered greatly. At the end of one segment of video, the feature points are detected once again. After the feature points of each video segment are obtained, direction, speed and location histograms of motion feature points are extracted and connected in series to serve as features of the video segments. Then, the features are subjected to Gaussian mixture modeling and updated in real time to obtain the probability of abnormal events in order to judge whether any abnormal event occurs or not. Thus, the abnormal events can be detected in real time.

The invention discloses a time sequence remote sensing image-oriented land utilization classification method, and the method comprises the following steps: carrying out principal component analysis onmultispectral images forming time sequence remote sensing data to obtain images of three principal components; pre-training each three-waveband image, and extracting a feature image; sequentially inputting the feature images into a semi-supervised convolution long-term and short-term memory network model for training according to a time sequence; and utilizing the trained model to predict and classify the image of the last time phase to obtain a classification result. According to the invention, time context information of a time sequence and space and spectral characteristic information of aremote sensing image are comprehensively considered; the problems existing in the prior art are well solved by utilizing the pre-training model and the semi-supervised classification learning mode, so the land utilization classification method is more suitable for classification scenes with less training sample data and missing or partially missing remote sensing image data, and a better land utilization classification result is obtained.