31results about How to "Fast Feature Extraction" patented technology

The invention relates to an intelligent terminal network teaching method. The intelligent terminal network teaching method comprises the steps that communication connection among the interior of a teaching terminal, the interior of a learning terminal, the interior of a parent management terminal, the interior of a server and the interior of a service platform is set up; the teaching terminal is used for recording the course of a teacher and uploading and storing the course of the teacher in the server; the learning terminal feeds information back to the teaching terminal through the server, and the teaching terminal answers the fed back information correspondingly; the parent management terminal is used for maintaining parent information; data are stored in the server and used for being called by the teaching terminal, the learning terminal and the parent management terminal, and the server is used for carrying out data interaction with other external devices; the teaching terminal and the learning terminal both are in information interaction with the service platform, a teacher management module, a student management module and an intelligent search module are arranged in the service platform, and the intelligent search module comprises an image retrieval module and an image recognition module; the learning module and a sign-in management module carry out information transmission. The intelligent terminal network teaching method can be used for developing a pre-class lesson preparing application, a classroom teaching application, an after school care application, a remote education application and the like based on the Internet.

Embodiments of the invention provide a big data feature extraction method and device. The device comprises a data structured module and a representative learning module, wherein the data structured module is used for pre-processing original big data and networking the pre-processed original big data to obtain a relationship network with nodes and an edges; and the representative learning module is used for obtaining high-dimensional vectors of the relationship network by adoption of an embedded mapping-based representative learning algorithm so as to obtain features of the original big data. The device provided by the embodiments of the invention can effectively extract the feature information in the big data without artificial participation; the feature information is uniformly expressed in a form of high-dimensional vectors, so that the features can provide a uniform effective processing method for a plurality of application services; and by adoption of the embedded mapping-based representative learning algorithm, structure information is retained in the high-dimensional vectors, so that more correct application services can be subsequently provided for the users.

The invention provides a vehicle license character feature extracting and classifying method using projection symmetry as the precondition for judgment. The method comprises the following steps: firstly extracting character feature to 26 letters and 10 numbers which may appear in vehicle license characters by using projection symmetry as precondition to divide into four classes, namely vertical projection class, horizontal projection class, central point projection class and dissymmetric property class and realize the coarse classifications of vehicle license characters; and fining, performing normalization transformation, then extracting features of points and rings and completing the fine classification of vehicle license characters. The invention combines the projection symmetry with the feature extraction method of points and rings to set a vehicle license character classifier, thus laying the foundation of realizing the identification of vehicle license characters finally. The method has better identification effect on confusable characters, such as '0' and 'D', '8' and 'B', '7' and 'T' and the like, thus increasing the identification speed and accuracy of vehicle license characters.

The invention provides a vibration type flow meter characteristic signal extraction method based on improved assemble average empirical mode decomposition. The method includes the steps of conducting end continuation processing on a collected vibration type flow meter flow vibration signal through a waveform-matched self-adaption end continuation method, conducting envelope line fitting on the collected vibration signal through a cubic B-spline method, conducting MEEMD decomposition to obtain a plurality of IMF components, conducting relevance analysis on the IMF components and the original signal, selecting the useful IMF components, conducting HHT conversion on the IMF components, and obtaining the Hilbert time-frequency spectrum and the marginal spectrum of the flow signal, wherein the Hilbert time-frequency spectrum and the marginal spectrum are the signal characteristics of the vibration type flow meter flow vibration signal. The method is suitable for accurately and rapidly metering the pipe network fluid flow in the industrial field.

The invention provides an SURF-based compression tracing method and an SURF-based compression tracing system. The method comprises the following steps of respectively collecting a positive sample and a negative sample of a t-th frame image, and positive samples and negative samples of the t-N-th to t-th frame images of a video file, using SURF to extract image features and a compression perception technology to obtain a first classifier and a second classifier, and weighting the first classifier and the second classifier to obtain a third classifier; using the third classifier to trace target images in the t+1 frame images of the video file. According to the target tracing method and system provided by the invention, due to the fact that the SURF extracted image features are adopted, the feature extracting speed is high, the image features can be still correctly extracted when a target environment is subjected to size change, rotation change and the like; original image features are compressed through a sparse random measurement matrix, so that the number of processed data is small, the updating speed of the classifiers is increased, and the target tracing method can be used for accurately tracing a target in a complex environment, and has real-time property and robustness.

The invention discloses a real-time gesture tracking method based on a cascade deep neural network. The method comprises: image data are obtained by a TOF camera and a color camera; an image preprocessor carries out preprocessing on the image data; essential characteristic extraction is carried out on the data after preprocessing by using a primary characteristic extractor; a cascade artificial neural network system carries out advanced characteristic abstracting processing; a mode matching device carries out mode matching according to the advanced abstract characteristics after characteristic abstract processing; and then an attitude processing center calculates all positions of twenty six nodes of a hand to obtain a hand attitude and spatial position data, and the data are transmitted to a computer application by a gesture attitude data flow. Rapid characteristic extraction, matching, and attitude calculation are carried out on image information of the human hand to guarantee the high stability, precision, and real-time performance of the calculated gesture attitude.

The invention provides a vehicle fine-grained classification method and device. The vehicle fine-grained classification method comprises the steps of extracting a basic feature map of an input picturethrough a convolutional neural network; adaptively constructing a global structure diagram according to the basic feature diagram; adopting a graph convolutional neural network to carry out association relationship reasoning on the global structure graph to form global guidance information; applying the global guidance information to the basic feature map to obtain an enhanced feature map; and inputting the enhanced feature map into a classifier for classification. According to the vehicle fine-grained image classification method, association relationship reasoning is carried out on differentparts of the vehicle, and the association relationship reasoning result is applied to the basic feature map, so that the obtained feature map has higher expression capability, and the vehicle fine-grained image classification accuracy is improved; in addition, the classification method does not need to mark key areas, and is quick in feature extraction, wide in application scene and low in calculation consumption.

The invention discloses a video duplicate checking method and system, a server and a storage medium. A video frame of a video to be subjected to duplicate checking is stamped at a specific time point,a feature value is extracted, if the feature value of the video frame stamped at the same time point as a comparison video are the same, it is determined that the two videos are the same, and due tothe fact that feature images needing to be stamped are small, the needed storage space is small, the feature value stamping speed is high, and the duplicate checking speed is high. Besides, the duplicate checking data is screened through basic information such as video length in the earlier stage of comparison; the data volume is reduced, an independent feature image is compared; a first feature image is converted into a first feature value, whether the feature values are the same or not is directly compared, when the Hamming distance between the first feature value and one of the second feature values is smaller than or equal to a first threshold value, it is considered that the first feature image is the same as a second feature image corresponding to a second feature value, XOR operation is only needed, complex encoding and decoding processing is not needed, and the duplicate checking speed is high.

The invention discloses a method and a system for effectively sharing and collecting. A method for storing websites comprises the following steps: classifying the stored websites when the websites are stored by a user; adding identifiers to the websites. The invention further discloses a method for recommending the websites. The method for recommending the websites comprises the following steps: analyzing using habits of the user; recommending the websites which are matched with the using habits of the user and contain specific classification and identifier characteristics to the user, or regulating the display priority of the websites of different classifications and identifiers according to the using habits of the user. By adopting the scheme of the invention, the characteristics of the websites can be extracted quickly, and the websites are classified according to the characteristics; the websites with certain characteristics can be recommended to the user according to the hobbies and interests of the user; the accuracy of websites recommending is improved; the interesting websites of the user are found by the user conveniently and more quickly.

The invention proposes a license plate character feature extraction and classification method which takes projection symmetry as a prerequisite judgment condition. In this method, the 26 letters and 10 numbers that may appear in the license plate characters are firstly extracted with projection symmetry as the judgment condition for character feature extraction, and they are divided into vertical projection symmetry, horizontal projection symmetry, central point symmetry, and no symmetry. Large categories, so as to realize the rough classification of license plate characters; after thinning and normalization transformation, the feature extraction of points and rings is carried out to complete the fine classification of license plate characters. The invention combines the projection symmetry with the point and ring feature extraction method to set the license plate character classifier, laying the foundation for the final realization of the license plate character recognition. This method has a good recognition effect on confusing characters, such as "0" and "D", "8" and "B", "7" and "T", which makes the recognition speed and accuracy of license plate characters get improved.

The invention provides a visual word bag model constructing model based on an improved SURF characteristic. A box filtering template added with gradient information is used to replace a Gaussian filter, and the template is more close to a Gaussian second-order differential template. In SURF characteristic expression, the overhead in time is reduced, and a SURF descriptor reduced to 32 dimensions while rotation invariance is ensured. When a word bag is constructed, an improved SURF algorithm is used to extract all improved SURF characteristic in an image library, a k-means clustering method is used to cluster all SURF characteristics as a visual word, and each image is expressed as the high-dimensional vector of the appearance frequency of each visual word. The comprises gradient information with rich images, one time of calculating a Haar wavelet is omitted, compared with the direct use of the SURF characteristic, the problem of non-uniform characteristic amounts extracted from different images can be solved well, the a word bag function allows multiple images to be expressed by a certain number of visual words, the space is saved, the processing is convenient, and the scalability is high.

The invention relates to an underwater image enhancement method based on an optimal restoration parameter. The method comprises steps that S1, images are inputted, hundreds of high quality underwaterpictures are collected as samples, and new underwater image dark channel apriority is acquired; S2, RGB three-channel transmission map estimation based on new UDCP conclusion is carried out; S3, backlight estimation based on a fusion method is carried out; and S4, image restoration is carried out. The method is advantaged in that the final output shows better enhancement in contrast, saturation and brightness, the method is suitable for the underwater images in different environments, and the enhanced images are applicable to underwater exploration, marine resource assessment and target identification applications.

The invention discloses a panoramic unmanned aerial vehicle image splicing method. The method comprises the steps of image input, image feature extraction, image feature matching, image matching, image splicing and spliced image output. The image input is an image shot by a plurality of unmanned aerial vehicles at the same time, and the image feature extraction utilizes an improved geometric algebra-scale invariant feature conversion algorithm an adaptive threshold algorithm is adopted for image feature matching, and a homography estimation algorithm which is robust through random sampling consistency is adopted for image matching. According to the method, an optimized geometric algebra scale invariant feature conversion algorithm is adopted to realize rapid feature extraction and featurematching; The limitation problems of large calculation amount and high splicing time cost are solved by applying a self-adaptive threshold method through large feature point extraction and splicing work; A random sample consistency method is adopted to estimate image transformation parameters, and a solution with the best consistency with data is determined. According to the image splicing method,the alignment speed of the images is greatly increased, and a satisfactory image splicing result is generated.

The SURF-based compression tracking method and system provided by the present invention use SURF to extract image features and compressive sensing by collecting the positive samples and negative samples of the t-th frame image and the t-N-th frame to the t-th frame image of the video file respectively. technology, obtain the first classifier and the second classifier, and weight the two to obtain the third classifier; use the third classifier to track the target image in the t+1th frame image of the video file. The target tracking method and system of the present invention, because SURF is used to extract image features, the feature extraction speed is fast, and the image features can still be correctly extracted when the target environment changes in scale, rotation, etc. The feature is compressed, so the processed data is less, and the update speed of the classifier is improved, so the target tracking method can accurately track the target in a complex environment, and has real-time and robustness.

The invention provides a dynamic vision sensor sample set modeling method based on a frame image, which can realize quick extraction of a moving target and has the advantages of low delay, low storagespace, high dynamic range and the like. The method comprises the following steps: S1, acquiring to-be-processed video data; s2, performing pixel normalization processing and pixel light intensity accumulation processing on continuous frame images in the to-be-processed video data; s3, performing unit discretization on the processed frame image to obtain a single pixel unit; s4, for each single pixel unit, performing calculating according to the light intensity change amplitudes of the current frame and the adjacent frame, and judging whether a trigger threshold T is reached or not; s5, storing the data information corresponding to the single pixel unit output in the step S4 according to a self-defined storage format, wherein the self-defined storage format comprises address event data corresponding to a single pixel unit; s6, coding all the stored data information according to actual requirements to form final sample set data.

The invention discloses an image comprehensive similarity analysis method based on description content and image content features. The method comprises the steps of extracting the image color information, namely the RGB model and image texture information, in image content feature information; extracting image title content and description content in the image description content feature information, and performing word segmentation processing; converting the RGB model into an HSV model, obtaining an HSV value of the image, and obtaining 24-dimensional feature information of the image color; calculating a gray value of the image in the image texture information; calculating cosine similarity to obtain image content similarity; obtaining features of the image title content and the image description content; calculating cosine similarity of the title contents and the description contents of the two images to obtain similarity of the description contents of the images; calculating image feature similarity; combining the image content features with the image description content features to generate composite features of the image; calculating cosine similarity of the image composite features to obtain similarity of the image composite features; judging whether images are similar or not only when the image feature similarity and the image composite feature similarity are both greater than or equal to the threshold.

The invention belongs to the technical field of information security, and discloses a video tampering detection method and system, a storage medium, a computer program and a terminal. The method comprises using a three-frame difference method to extract a video key frame; carrying out DCT on the obtained key frame to reduce high-dimensional video data and carry out de-noising processing to obtaina sub-graph containing most signal energy; performing feature extraction on the sub-graph through an ORB algorithm; and comparing the similarity between a to-be-detected video and an original video through the Hamming distance. The detection performance of the proposed method is evaluated by using a statistical method, feature extraction is performed on the key frame of a test video by using an SIFT algorithm, an ORB algorithm and the algorithm of the invention, and the average feature point number, the average time consumption and the average storage space of the key frame are counted. According to the embodiment of the invention, the number of feature points extracted from the key frame is far less than the number of feature points extracted by an SIFT algorithm and is 40% less than thenumber of feature points extracted by an ORB algorithm.

The invention provides a multi-modal emotion recognition method based on acoustic and text features, which is suitable for extracting speech and text emotion features. The method comprises the following steps: extracting emotional shallow-layer features of input voice by using OpenSMILE, and fusing the emotional shallow-layer features with deep-layer features obtained by learning the shallow-layer features through a Transform network to generate multi-level acoustic features; performing forced alignment on the voice and the transcriptional text to obtain pause information, encoding the speaking pause information in the voice, adding the encoded speaking pause information to the transcriptional text, sending the encoded speaking pause information to a hierarchical dense connection DC-BERT model to obtain text features, and fusing the text features with acoustic features; according to the method, effective context information is acquired by utilizing priori knowledge through a BiLSTM network, a part which highlights emotion information in features is extracted through an attention mechanism to avoid information redundancy, a global average pooling layer is added behind the attention mechanism to replace a traditionally used full connection layer, and finally the information is sent to a softmax layer for emotion classification. The method has the advantages of simple steps, accurate identification and wide practical value.

The invention discloses an image comprehensive similarity analysis method based on description content and image content features, comprising: extracting image color information in image content feature information, that is, RGB model and image texture information; extracting image title in image description content feature information Content and description content, word segmentation processing; convert the RGB model to the HSV model, obtain the HSV value of the image, and obtain the 24-dimensional feature information of the image color; calculate the gray value of the image in the image texture information; calculate the cosine similarity, and get Image content similarity; obtain the characteristics of image title content and image description content; calculate the cosine similarity between the title content and description content of two images, and obtain the similarity of image description content; calculate image feature similarity; Combined with image description content features to generate image composite features; calculate the cosine similarity of image composite features to obtain the similarity of image composite features; when both image feature similarity and image composite feature similarity are greater than or equal to the threshold, the image is judged to be similar .

The visual bag-of-words model construction method based on improved SURF features uses the box filter template with gradient information instead of Gaussian filter, which is closer to the Gaussian second-order differential template; when expressing SURF features, it reduces the time overhead, and in While ensuring the invariance of rotation, the SURF descriptor is reduced to 32 dimensions; when constructing the bag of words, the above-mentioned improved SURF algorithm is used to extract all the improved SURF features in the image library, and the k-means clustering method is used to cluster all the SURF features into visual words , so that each image is represented as a high-dimensional vector of the frequency of each visual word. This method contains more abundant gradient information of the image, and omits a Haar wavelet calculation step; compared with directly using SURF features, it can well solve the problem that the number of features extracted from different images is not uniform, and the bag-of-words model can Represent multiple images with a certain amount of visual words, which saves space, is convenient to process, and has strong scalability.

The invention discloses a bearing fault detection method based on a variable learning rate multi-layer perceptron. The method comprises the following steps: Step 1, collecting and extracting a vibration signal of a bearing by using a sensor; step 2, carrying out feature iteration on the vibration signal by using a multilayer perceptron; 3, a back propagation process is used for automatically optimizing the weight coefficient of the multi-layer perceptron, the learning rate of automatic optimization is changed according to the number of iterations, time domain signals obtained after iteration enter a support vector machine classifier, and a diagnosis model is obtained; step 4, inputting test data into the diagnosis model for diagnosis, and completing the detection of the bearing fault; according to the method, the bearing signal faults are extracted in a crossed mode through the multi-layer perceptron based on the variable learning rate, the loss function value can be rapidly reduced to complete fault diagnosis, the feature extraction speed is high, data do not need to be preprocessed, diagnosis precision is high, and the method has the advantages of being rapid in feature extraction, high in precision and good in robustness.

The invention provides a multi-user barrier-free code scanning payment method applied to an IOS system. A self-learning-based machine learning model, two full screens, two cameras and a mobile paymentdevice which is based on an IOS control system of a central processing unit and supports cooperative application of an input device and a printing device are configured in multiple scenes for networking, and the mobile payment device is connected with an electronic terminal and a server cluster; obtaining parameter data to generate a two-dimensional code image; when it is monitored that the payment event is triggered, code scanning and non-contact card swiping payment are supported, and the first distance measurement sampling point is monitored, self-learning image reading and payment of multiple users are completed through deduction scanning. According to the method, a preset algorithm is adopted for decoding, and when a bar code exists in front, a display screen is changed into transparent glass for reading; through algorithm processing, bar code content and display information are simultaneously read based on the two full screens and the two cameras. The invention further providesa multi-user barrier-free code scanning payment device applied to the IOS system.

The present invention is applicable to the technical field of video monitoring, and provides a fast feature extraction system based on massive video, including: massive video analysis depth application module A, massive video analysis application service module B, massive video source module C, and video big data processing Module D and Basic Resource Module E. The video big data processing module D includes an algorithm control module F, a task segmentation module G, an algorithm injection module H, a task scheduling module I and a data storage K. After receiving the job task, the video big data processing module D calls the task segmentation module G according to the video source information to segment the job and package it into small tasks; through the task scheduling module I, the small tasks are sent to each cluster analysis node, and the analysis node The video slice is analyzed through the algorithm control module F and the algorithm injection module H; the cost and bandwidth requirements are greatly reduced by deploying multiple relatively low-performance servers for distributed processing.

The invention provides an internet big data analysis and extraction method, which comprises the following steps of: 1, dividing a data object into different parts and types according to the characteristics of data to obtain a data range to be extracted; step 2, establishing a regression model, solving each parameter of the model according to the measured data, then evaluating whether the regression model can fit the measured data, and if the regression model can fit the measured data, further narrowing the range of the data to be extracted according to the independent variable; step 3, dividing the data into more than two aggregation classes according to the characteristic attributes of the data, and grouping the data to be grabbed, the elements in each aggregation class having the same characteristics; 4, calculating the similarity degree of the two pieces of data by adopting a similarity matching method; 5, the word frequency is used as a statistical index to indicate data segment information fed back by the data; and step 6, obtaining a data analysis result. The method is automatically completed by using an embedded mapping-based representation learning algorithm, and the calculation efficiency is high.

The invention provides a method for reading a two-dimensional code under a screen based on a photoelectric sensor, and the method comprises the steps: carrying out the networking of a mobile payment device which is arranged in a plurality of scenes and is provided with the photoelectric sensor, a camera and a control system based on a central processing unit and supports the cooperative application of an input device and a printing device, and enabling the mobile payment device to be connected with an electronic terminal and a server cluster; obtaining parameter data to generate a two-dimensional code image; when it is monitored that the payment event is triggered, code scanning and non-contact card swiping payment are supported, and the first distance measurement sampling point is monitored, reading is completed through deduction scanning. According to the method, for a two-dimensional code image, a preset algorithm is adopted for decoding, when various devices sense that a bar code exists in front, display content is removed, and a display screen is changed into transparent glass to facilitate reading; through algorithm processing, the bar code content is read while the content is displayed on the screen. And the flexibility and usability of code scanning are improved. The invention further provides a device for reading the two-dimensional code under the screen based on the photoelectric sensor.

The invention discloses a rolling bearing fault detection method based on a multilayer residual network model. The rolling bearing fault detection method comprises the following steps: Step 1, collecting and extracting a vibration signal of a bearing by using a sensor; step 2, carrying out feature iteration on the vibration signal by utilizing a residual network; 3, the time domain signals subjected to iteration enter a neural network full-connection layer, weighted summation is conducted on the time domain signals through the network full-connection layer, the weight coefficient of a network structure is automatically optimized through the back propagation process, and a diagnosis model is obtained; step 4, inputting test data into the diagnosis model for diagnosis, and completing the detection of the bearing fault; according to the method, bearing signal faults are extracted in a cross mode based on the multi-layer residual error network model, so that the loss function value can be rapidly reduced to complete fault diagnosis, the feature extraction speed is high, data does not need to be preprocessed, the diagnosis precision is high, and the method has the advantages of being rapid in feature extraction, high in precision and good in robustness.

The present invention provides a dynamic visual sensor sample set modeling method based on frame images, which can realize fast extraction of moving targets, and has the advantages of low delay, low storage space, high dynamic range and the like. It includes: S1: Acquire the video data to be processed; S2: Perform pixel normalization processing and pixel light intensity accumulation processing on the continuous frame images in the video data to be processed; S3: Perform unit discretization on the processed frame images to obtain Single pixel unit; S4: For each single pixel unit, calculate according to the light intensity change range between the current frame and the adjacent frame, and judge whether the trigger threshold T is reached; S5: According to the custom storage format, store the output in step S4 Data information corresponding to a single pixel unit; custom storage formats include: address event data corresponding to a single pixel unit; S6: All stored data information is further encoded according to actual needs to form the final sample set data.

The present invention relates to an underwater image enhancement method based on optimal restoration parameters, the method includes the following steps: the method includes the following steps: Step S1, input images, collect hundreds of high-quality underwater pictures as samples, Obtain the dark channel prior of the new underwater image; step S2, estimate the RGB three-channel transmission map based on the new UDCP conclusion; step S3, estimate the background light based on the fusion method; step S4, restore the image. Its advantages are as follows: the final output results show better enhancement effects in many aspects such as contrast, saturation, and brightness. Moreover, the method of the present invention is applicable to underwater images in different environments, and the enhanced images can be used in applications such as underwater detection, marine resource assessment, and target recognition.