461results about How to "Avoid double counting" patented technology

The invention relates to a statistical device and a detection method for television audience ratings based on human face detection, and the device includes a dual camera, an infrared lamp, an infrared lamp light filtering device, a set-top box and a background server, wherein, the dual camera, the infrared lamp and the infrared lamp light filtering device are in a parallel connection, the dual camera is connected with another set-top box, and the set-top box is connected with the background server in a wired or a wireless mode, and the set-top box is also connected with the television set. The invention has the advantages that identity and watching state are recognized and confirmed by biological characteristics; number of people for watching specific advertisement in each time period, preference degree as well as age, gender, level of living of watcher are counted, detected human face is in real-time tracking, thereby avoiding multiple counting and improving accuracy of television audience rating.

The invention provides a method for quickly invoking program instructions, comprising that coding character string input by the user is received; in a prearranged mapping table, a triggering condition is queried according to the coding character string or the content which is converted by the coding character string; if records complying with the matching exist, the instruction is started to trigger an interface and unfold corresponding information; the triggering information of the user is received and corresponding instruction invoking information is gained; the instruction invoking information is sent to corresponding interfaces to realize the invoking of corresponding program instructions. As the technical proposal of the invention is based on keyboard input, the instruction inlet is not needed to be searched in the screen, thus avoiding the problems of low efficiency of positioning means such as menu, button, icon, etc., which are based on graphic interfaces; secondly, the technical proposal of the invention can support the operation system instruction and application program instruction at the same time, thus unifying the inlet of computer program instructions and improving the efficiency of the instruction invoking.

In order to improve the efficiency of vehicle positioning to meet the real-time requirements for vehicle type identification, the present invention provides a vehicle type identification method and system based on a deep neural network. The method comprises: integrating candidate target extraction and target identification into a network, and using end-to-end detection / identification methods to integrate feature extraction, target location, and target detection into a single network. According to the method and system provided by the present invention, target extraction t is no longer extracted from the original image, but is extracted from the high-dimensional feature map with very small dimensions by using the reference point and multi-dimensional coverage manner, and at the same time, the method for sharing the deep convolutional network parameters is used in the feature extraction layer, so that repetitive calculation of features is avoided, the identification efficiency is greatlyimproved, 20fps is reached in reality, the effect of the processing efficiency of a single server reaches 2 million sheets / day, and the requirement for real-time vehicle type identification is satisfied.

The invention discloses a tidal current optimization method of an electric power system based on partial automatic differentiation technique. Compared with the existing tidal current optimization methods based on the automatic differentiation technique, the algorithm fully utilizes the unchangeable characteristic of most elements in a Jacobian matrix and / or a Hessian matrix of an objective function and a constraint function during iteration, and is added with a function of identifying invariable elements in the Jacobian matrix and the Hessian matrix, and stores the invariable elements in a list before the first iteration; in each iteration of a numerical optimization algorithm, variable elements in the Jacobian matrix and / or the Hessian matrix can be computed only by the automatic differentiation technique. The tidal current optimization method based on the partial automatic differentiation technique can greatly reduce the burden on software developers and maintainers, improve the maintainability and the flexibility of a tidal current optimization application program, efficiently support customized models, and satisfy the analysis, running and scheduling requirements of the modern power system on the premise that the computational efficiency is not reduced substantially.

The invention discloses a cloud platform-based large-scale centralized remote wind electric power prediction center, which is built in a wind electric power prediction system on an enterprise-level personnel cloud platform, and comprises a cloud real-time historical database system, a relational database, a downscaling large-scale parallel mode computation module for numerical weather prediction, a model computation module, a statistic analysis module, a data input / output interface and a man-machine interaction interface. In the prediction center, the traditional wind electric power prediction system is transferred onto the enterprise-level personnel cloud platform, the downscaling large-scale parallel mode computation module for numerical weather prediction is additionally arranged, so services can be provided for multiple wind farms through the strong and flexible computation capability and storage capability. While the safety is ensured, the maximum resource intensification, rapid implementation of system deployment and the minimum expenses of maintenance cost are realized, and the prediction center can be suitable for the severe environments of future offshore wind farms.

The invention discloses a haze image rapid haze removal method. Dark channel rapid computation comprises the steps of establishing a cross-shaped dark channel formwork; computing pixel values of a transverse dark channel window and a longitudinal dark channel window; and comparing the pixel values and assigning a minimum value to an image dark channel. A mean value of a maximum gray value of 0.1% of a pixel in the dark channel is selected and serves as an atmosphere light value; a transmission image is obtained according to a dark channel image; and the transmission image is refined through a rapid Gaussian filter. Due to the fact that the cross-shaped formwork is adopted, computing area and computing amount are reduced; repeated computation is avoided and computing time is saved; the atmosphere light value is computed and obtained in the dark channel in the process of estimation and computing speed is accelerated under the premise of guaranteeing haze removal effect; and a transmission rate is optimized through a recursion Gaussian filter method and an optimized result is good.

This invention discloses one cross analysis process system multi-dimension expression data buffer method and device, which comprises the following steps: setting in advance multi-dimension buffer labels in the expression sentences; analyzing sentence to generate grammar tree and executing the expressions; in process, if yes, then setting the buffer label property as true and getting the said buffer label computation results and generating relative key words; storing the computation results and said key words.

The embodiment of the invention discloses a visitor flow rate statistical method and device based on image recognition, and the method comprises the steps: sequentially obtaining target human faces in N target images, so as to obtain a first target human face set, wherein N is a positive integer; carrying out the filtering of the first target human face set based on a human face tracking algorithm, so as to obtain a second target human face set; carrying out the filtering of the second target human face set based on a human face feature value real-time clustering algorithm, so as to obtain a third target human face set; and counting the target human faces in the third target human face set, so as to obtain a visitor flow rate statistical value. The method carries out the two-layer filtering of N target images, and can accurately filter out repeated human faces in the target human faces, so as to avoid the repeated counting of target human faces, and improve the visitor flow rate statistical accuracy.

The present invention discloses a distributed OLAP analysis method and system based on pre-computation. The method mainly comprises: constructing a data model based on the distributed data warehouse, and defining data cubes according to the data model; starting pre-computation tasks for the given data cubes so as to pre-build the cubes in a parallel computation manner, and storing results in a distributed key storage system; through a series of steps, converting multi-dimensional analysis operations into key-value query operations on the data cubes, directing obtaining analysis results from the built cubes, and displaying the results in the form of rich and diversified charts; and using the NoSQL to carry out cache optimization on the OLAP query operation. According to the method and system disclosed by the present invention, the powerful processing performance of the Hadoop platform is fully exerted, and data cubes are pre-built, the problem that the query is slow due to that the traditional method needs a large amount of computation from the original data in each query is overcome, so that OLAP analysis efficiency and system performance are improved.

The invention relates to a network intrusion detection method based on association rule classification. The network intrusion detection method comprises the steps of pre-processing network data, extracting an association rule, classifying network connection data and displaying a classification result. According to the invention, on the basis of an improved Apriori algorithm (Apriori-index), a KDDCup99 network connection data set, namely an international standard data set, is taken for example; firstly, the association rule is extracted from network connection data selected from the KDDCup99 network connection data set; then, test network connection data is classified according to the association rule; therefore, whether current network connection is attack connection or not can be judged; the specific attack type of the current network connection can also be judged; and related statistical data is displayed. The Apriori-index algorithm is more suitable for the KDDCup99 data set; the association rule extraction speed and the network connection classification speed are greatly increased; the accuracy of a detection result is also improved; and the disadvantages of slow classification speed and high false alarm rate in the traditional intrusion detection system are improved to a certain degree.

The invention discloses a method for detecting ampoule bottle appearance quality through a medical visual inspection robot. The method comprises the steps of (1) acquiring an ampoule bottle image, and preprocessing the image through a median filter; (2) solving an ampoule bottle mimic symmetry line by employing a threading method to obtain a rotating angle alpha; (3) rotating the image by using a bilinear interpolation-summation method; (4) solving a symmetry center line, wherein the intersection of the symmetry center line and the head is a symmetry center point h0; (5) solving a head edge point; (6) enabling the outer curve at the head of the ampoule bottle to have the characteristic of a symmetrical variable trend in a lower parabola shape, and judging whether the ampoule bottle appearance wire-drawing quality is qualified; (7) positioning a black head detection rectangular area; (8) judging whether the black head ampoule bottle appearance is qualified by employing a difference method; and (9) integrating the appearance drawing and black head quality detection result. According to the method, the problems that the manual detection efficiency is low, the consistency is low and the leakage detection rate is high and the like are overcome, and the degree of automation and product quality of the ampoule bottle medicine production are improved.

The embodiment of the invention discloses an internet advertisement anti-spamming method and an internet advertisement anti-spamming system, so as to solve the problem that the statistical result of information, such as impressions or click rate of an internet advertisement is inaccurate caused by incapability of distinguishing spamming log data from non-spamming log data due to the fact that the spamming log data is mingled with all log data. The method comprises the steps of reading and partitioning various log data through a first single-thread; counting data required by various predetermined spamming rules in the log data partitioned by the first single-thread through a first multi-thread; reading and partitioning various log data through a second single-thread; judging whether the log data partitioned by the second single-thread is spamming log data or not through a second multi-thread according to the predetermined spamming rules and the data required by the predetermined spamming rules; and marking the log data partitioned by the second single-thread according to a judging result. The log data is only required to be counted for one time, so that the repeated counting of log data in the same dimensionality is avoided.

The invention discloses an incremental learning and face recognition method based on locality preserving nonnegative matrix factorization (LPNMF), and relates to the technical field of pattern recognition. The incremental learning and face recognition method is essentially a face recognition method based on on-line learning. The method comprises the step (a) of face image preprocessing, the step (b) of initial sample training, the step (c) of incremental learning and the step (d) of face recognition. The incremental learning and face recognition method can be applied to a linear face recognition system and a non-linear face recognition system, keeps local structures of original space of face images, greatly improves recognition rate, and can be actually applied in an on-line mode.

The invention discloses a Canny edge detection algorithm based on a local threshold, and the algorithm mainly comprises the following steps: 1), carrying out the smooth filtering of an image, and removing noise through Gaussian filtering; 2) calculating an image gradient amplitude and direction based on the smoothed image; 3) performing non-maximum suppression calculation on the gradient directionto obtain an edge point, if the amplitude of the gradient of the current pixel point is greater than the gradient amplitude of two adjacent pixel points in the positive and negative directions of thegradient, regarding the point as the edge point, and marking the corresponding position of the point as 1, otherwise, suppressing the point as a non-edge point, and marking the corresponding positionof the point as 0; 4) obtaining a global proportion value; 5) calculating a local high-low threshold value matrix through an acceleration algorithm according to the global proportion value, and 6) detecting edges according to a double-threshold value matrix to obtain a final edge image. The Canny algorithm can detect the local significant edges in the image in a self-adaptive mode and has good acceleration calculation performance.

The invention discloses a method and equipment for common detection of human faces and human bodies. The method comprises the steps that standard data is obtained, wherein the standard data comprises position frame information for labeling the positions of the human faces and the human bodies of all pedestrians; a common recognition model is corrected through the position frame information in the standard data, wherein the common recognition model is generated based on Faster RCNN; common detection of the human faces and the human bodies is carried out on images in a video to be recognized through the corrected common recognition model to output real-time structuralization information, wherein the structuralization information comprises human face information and human body information of the pedestrians at the same time. Thus, efficient simultaneous recognition is achieved, resources are saved, practicality is improved, and the corresponding relation between the human faces and the human bodies is guaranteed.

The invention discloses a method for positioning a plane centroid of a pipe orifice image of a condenser. The method comprises the following steps of: (1) calibrating a camera; (2) controlling a hand-eye mechanism of a robot to move along an optical axis of the camera, continuously shooting gray aerial views of a pipe plate area to be cleaned, selecting a clear image from the gray aerial views, pre-processing the clear image to obtain a plurality of binary images of which pipe orifices and backgrounds are in black-white contrast, and distributing the pipe orifices in the pipe plate area to becleaned in a horizontal and longitudinal staggered mode; (3) drawing a column pixel gray value accumulated sum curve S1 and a line pixel gray value accumulated sum curve S2, counting horizontal coordinates of a local minimal value point set at a column trough in the S1 and longitudinal coordinates of the local minimal value point set at a line trough in the S2, and thus obtaining a centroid roughcoordinate set; and (4) eliminating falsified centroids to roughly position the centroid in a multi-target mode. The method has the advantages of high accuracy and small calculation amount.

The method includes following steps: the kernel density estimation function is used to build a modeling for pixel process; the distribute free probability density is used to estimate the distribution of pixel gray scale of image calculated by theory. The method gets the sample set of time domain diversity from original training sequence to use it in modeling training. In the procedure of extracting background and testing moving objects, the storage and usage of original training data is not needed, that saves storage space and avoids multiple calculations.

The invention provides an end-to-end tendency scene character detection and recognition method and an end-to-end tendency scene character detection and recognition system. The method comprises the steps: collecting pictures marked with object types, constructing a training data set and a test data set, and carrying out the preprocessing of the data sets for training a feature extraction network; collecting pictures marked with character positions, constructing a training data set and a test data set, and preprocessing the data sets to train a network of a character detection part; pictures marked with character positions and character contents at the same time are collected, a training data set and a test data set are constructed, and corresponding picture preprocessing is carried out to train a neural network of a character recognition part; inputting the picture into a convolutional neural network; extracting a shared convolution feature map through a feature extraction network; andmapping predicted character position coordinates to the shared convolution feature map by the character detection part maps by using the shared convolution feature map, cutting off a feature map blockcorresponding to the character part in the image and converting the feature map block into a feature sequence; finally decoding the feature sequence into a readable character sequence. Since the convolution characteristic spectrum is only calculated once, the intermediate redundancy process is avoided. The speed of the whole scene character detection and recognition system is improved.

The invention discloses a recommendation system recall method based on an attention mechanism, and the method comprises the following steps: extracting user features and commodity features in a training sample, converting the user features into a user embedding vector, and converting the commodity features into a commodity embedding vector; inputting the user embedding vector and the commodity embedding vector into an attention mechanism model for training, learning the weight of each feature through an attention network in the model, and performing weighted summation on the embedding vectorsof all the features according to the weights to obtain a user representation vector and a commodity representation vector; calculating an inner product of the user representation vector and the commodity representation vector to obtain a user commodity purchase intention matching degree of the training sample, establishing a cross entropy loss function of the user commodity purchase intention matching degree, calculating a minimized cross entropy loss function, and converging an attention mechanism model; and inputting the to-be-tested sample into the converged attention mechanism model, obtaining a user commodity purchase intention matching degree of the to-be-tested sample, and selecting a commodity of which the user commodity purchase intention matching degree is in a preset interval asa recall result for recommendation. According to the invention, the generalization is enhanced, and the calculation amount of recall recommendation is greatly simplified.

The invention discloses an implementation method and an implementation device for data duplication elimination query. The implementation method comprises the following steps: querying all database nodes in a plurality of database nodes to obtain query results, wherein for all the database nodes querying the query results, the query results queried by the database nodes are subjected to the duplication elimination operation, and the query results subjected to the duplication elimination operation are taken as the query results of the database nodes; merging the query results of the database nodes. The implementation method and the implementation device achieve the duplication elimination query operation of large-scale data in a database cluster, avoid the condition that duplication elimination query operation is confined to the duplication elimination division column, achieve the duplication elimination query to any data columns, and in addition, avoid the problem of repeated calculation in duplication elimination query, and improve the duplication elimination query efficiency.

The invention relates to a method for emitting detector which can select the detected target via swing-by mechanism, belonging to the technique of deep space detecting rail transfer. It can solve the problems of approachability evaluation of detected target with long semi long axis or large eccentric rate. The inventive method comprises: calculating optimized dual impulse transfer track; selecting swing-by track type; and assembling the swing-by tracks. First calculating the optimized dual impulse transfer track from the target star to emitting star; using the emitting star as swing-by star; using the parameter of dual impulse transfer at the emitting star as the match parameter in swing-by flying; using the heliocentric large-ellipse orbit whose period is several times of the revolution period of emitting star and the deep space device at aphelion, to search for the emitting parameter of said match condition. The invention guides in the swing-by mechanism to realize the expansion of former dual impulse transfer, to improve the approachability of the target with long semi long axis or large eccentric rate.

The invention discloses a high-temperature forged part surface defect in-position detecting method based on PCA and CNN. The method comprises the steps of acquiring a three-dimensional point cloud data set of a to-be-detected finished forged part through a three-dimensional laser scanner according to a preset sequence rule, and calculating a main curvature for obtaining a five-dimensional point cloud data set of the finished forged part; reducing the number of dimensions of the five-dimensional point cloud data set of the finished forged part to two by means of a principal component analysis (PCA) method, and inputting the two-dimensional point cloud data into a CNN forged part surface defect detector, and outputting a determining result by the CNN forged part surface defect detector. The CNN forged part surface defect detector is trained through using the point cloud data of a standard forged part, and furthermore the successively trained CNN forged part surface defect detector is utilized for performing in-position detection for determining whether a defect exists on the surface of the forged part with high temperature above 1000 DEG C, namely the structure defects such as corner missing, projection or recess, thereby realizing in-position quick and accurate detection to the high-temperature forged part, and ensuring high production quality of the forged part.

The embodiment of the invention discloses a speed forecasting method which comprises the following steps: calculating section lengths of all planned sections in a processing path and maximum forecasting speed of section end points corresponding to the section lengths; saving the section lengths of all the planned sections and the calculation results of the maximum forecasting speed of the section end points corresponding to the section lengths to finish the preprocessing phase; and performing forecasting processing to each planned section according to the calculation results. The embodiment of the invention further discloses a numerical control device and a numerical control system. By virtue of the methods, the safety and the reliability for processing of the numerical control machine tool is ensured, and the processing efficiency and the processing quality are improved.

The invention discloses a cardinality estimation method aiming at streaming big data. The method is characterized in that cardinality estimation efficiency is increased by lowering calculation precision; partition calculation is performed on the intermediate statistical information needed by as HyperLogLog Counting algorithm, an efficient Hash algorithm and an optimal bucketing number are selected, an improved bucketing method is used to evenly map hashed data into different bucket numbers, increment maintenance is performed, and combination is then performed to obtain the final intermediate statistical information so as to calculate a cardinality estimation value. The method has the advantages that historical data is effectively utilized, repeated calculation is avoided, and the cardinality estimation efficiency is increased greatly; high-precision cardinality estimation is achieved, and the efficient bucketing method is provided as compared with a traditional algorithm; the algorithm is quite low in space complexity, and calculation resource consumption is lowered.

The embodiment of the invention discloses a method for building a user image. The method includes the steps that an original behavior log of the user is obtained; the original behavior log is preprocessed to generate a target behavior log of the user; the target behavior log is analyzed to generate the latest behavior sequence of the user; a historical behavior sequence of the user is called; the latest behavior sequence and the historical behavior sequence are fused, and an attribute label of the user is obtained through calculation; the user image of the user is built according to the attribute label. By using the historical behavior sequence of the user, repeated calculation can be avoided in the process of user behavior sequence generating which consumes most resources, then hardware resources are saved, and the building efficiency of the user image is improved. Correspondingly, the embodiment of the invention further discloses a device for building the user image.

The invention discloses a permanent magnet synchronous motor model prediction control method comprising the following steps: (1) converting an acquired three-phase current to an alpha-beta coordinate system at a current moment k; (2) predicting the current value at a moment k+1 and a current reference value at a moment k+2 according to the current value of the alpha-beta coordinate system; (3) calculating the voltage control vector reference value at the moment k+1 according to the current value at the moment k+1 by taking the current reference value at the moment k+2 as a control objective; (4) analyzing out the angle corresponding to the voltage control vector reference value at the moment k+1, determining a sector to which the angle belongs, and extracting candidate voltage control vectors from the sector; and (5) substituting the candidate voltage control vectors into an objective function, and selecting the candidate voltage control vector corresponding to the minimum objective function value as an optimal voltage control vector. A lot of repeated calculation is avoided, the voltage vector is more accurate, the torque ripple of motors is reduced, and delay of the microprocessor due to calculation is offset.

The invention discloses a traffic flow visual inspection method. Traffic scene image sequences are acquired through electronic monitoring in real time, a traffic scene weather condition environment is sensed by an illumination intensity sensor, a temperature sensor and a humidity sensor, daytime / night, fine day / rainy day / normal day and other weather conditions are judged, and illumination and shadow pretreatment is performed correspondingly. Running distances are restrained according to traffic rules, meanwhile lane departure phenomenon caused by overtaking and passing-by in the actual vehicle running is considered, double virtual lines are arranged at the same horizontal positions of all of lanes within a monitoring range in images, vehicle positions are quickly detected and positioned by utilizing double-template matching convolution in a double virtual line detecting region, 'one-to-more' and 'more-to-one' phenomena are eliminated, and wrong detection and wrong judgment are decreased. Intervals of vehicles in the horizontal direction and the vertical direction are judged and identified, vehicle target positions are restrained according to the horizontal and vertical position information of the vehicles and correctly positioned, normally running vehicles are counted, traffic flow statistic is performed, and the problem of inaccurate traffic flow counting is solved. The traffic flow visual inspection method has high detecting accuracy and good anti-interference performance and real-timeliness.

The invention provides a passenger flow counting method based on deep learning in a vertical visual angle. According to the method, passenger flow statistics is performed in the vertical visual angle. Compared with an oblique photographing visual angle, the visual angle realizes easier coping with conditions such as market, supermarket and subway with high passenger density. The method provides head and shoulder detection by means of deep learning. Strong learning capability of deep learning is utilized. Background modeling and foreground extraction on a video are not required, and passenger cutting on the foreground is not required. Head information and shoulder information can be more accurately detected with higher robustness. A deep convolutional characteristic is utilized for performing matched tracking. Compared with manually designed characteristics such as HOG and LBP, the deep convolutional characteristic has better expression capability and can better cope with various scenes. The passenger flow counting method directly use the deep convolutional characteristic of a certain layer for matching, thereby preventing repeated characteristic calculation and realizing high time saving effect.

The invention discloses a wireless sensor network (WSN)-technology-based indoor three-dimensional accurate positioning method. WSN indoor three-dimensional accurate positioning is realized by the following steps of: 1) deploying sensor network nodes with known locations in a given area; 2) planning the sensor network nodes with the known locations in a three-dimensional relative coordinate system; 3) establishing a three-dimensional Hash table indicating correspondence between signal intensity and coordinate values in advance, and generating a table lookup KEY value; 4) generating a simple three-dimensional route map for the map of a given area according to possible routes; 5) synchronizing the time of each node by using a wireless sensor network time synchronization algorithm; 6) determining the three-dimensional locations of mobile nodes by continuously exchanging information between the continuously mobile sensor nodes and deployed fixed sensor network anchor nodes; and 7) receiving anchor node signals for many times, and looking up the three-dimensional Hash table to obtain the coordinate values for data fusion. Three-dimensional positioning results are continuously corrected on the three-dimensional route map by continuous circular testing to obtain accurate location coordinates.

The invention discloses a space time code three-dimensional self-adaptation clutter cancelling method for onboard multiple input multiple output (MIMO) radar. The method includes first building an onboard MIMO radar echo signal model, calculating residue echo signal energy, minimizing the echo signal energy to obtain a self-adaptation clutter canceller coefficient compensation device, acquiring a space time code three-dimensional self-adaptation clutter canceller through sliding window arrangement, utilizing the canceller to cancel echo signals and cancelling most clutter through preprocessing of the space time code three-dimensional self-adaptation clutter canceller. The method is used before a common space time code processing method and a dimension reducing STAP method to conduct space time code three-dimensional canceling on the clutter in the onboard MIMO radar echo signals, the clutter can be effectively restrained in the first stage of radar signal processing, and detection performance of a moving target is improved. The method is applicable to both positive side looking radars and non-positive side looking radars, and has good moving target detection capability.