102results about How to "Reduce the amount of calculation data" patented technology

The invention relates to a multi-scale normal feature point cloud registering method. The multi-scale normal feature point cloud registering method is characterized by including the steps that two-visual-angle point clouds, including the target point clouds and the source point clouds, collected by a point cloud obtaining device are read in; the curvature of radius neighborhoods of three scales of points is calculated, and key points are extracted from the target point clouds and the source point clouds according to a target function; the normal vector angular deviation and the curvature of the key points in the radius neighborhoods of the different scales are calculated and serve as feature components, feature descriptors of the key points are formed, and a target point cloud key point feature vector set and a source point cloud key point feature vector set are accordingly obtained; according to the similarity level of the feature descriptors of the key points, the corresponding relations between the target point cloud key points and the source point cloud key points are preliminarily determined; the wrong corresponding relations are eliminated, and the accurate corresponding relations are obtained; the obtained accurate corresponding relations are simplified with the clustering method, and the evenly-distributed corresponding relations are obtained; singular value decomposition is carried out on the final corresponding relations to obtain a rigid body transformation matrix.

The invention discloses a real-time image semantic segmentation method based on a lightweight full convolutional neural network. The method comprises the following steps of 1) constructing a full convolutional neural network by using the design elements of a lightweight neural network, wherein the network totally comprises three stages of a feature extension stage, a feature processing stage and acomprehensive prediction stage, and the feature processing stage uses a multi-receptive field feature fusion structure, a multi-size convolutional fusion structure and a receptive field amplificationstructure; 2) at a training stage, training the network by using a semantic segmentation data set, using a cross entropy function as a loss function, using an Adam algorithm as a parameter optimization algorithm, and using an online difficult sample retraining strategy in the process; and 3) at a test stage, inputting the test image into the network to obtain a semantic segmentation result. According to the present invention, the high-precision real-time semantic segmentation method suitable for running on a mobile terminal platform is obtained by adjusting a network structure and adapting asemantic segmentation task while controlling the scale of the model.

The invention provides a method for quickly recognizing a speaker and belongs to speaker recognition methods. The method comprises the following steps of: combining a Gaussian mixture model, and taking the supervector of the Gaussian mixture model as the feature parameter of the speaker; taking the supervector of the Gaussian mixture model as input, designing a one-class support vector machine classifier; and training N classifiers corresponding to N speakers, thus obtaining a voice sample of one speaker from one classifier. By utilizing the method, the speaker recognition speed is increased; for every new registered speaker, only one one-class support vector machine classifier is trained for the new speaker, so that the speaker recognition system has good extensibility.

The invention belongs to the field of traffic condition prediction, and relates to a method for discovering congestion points, congestion lines and congestion areas based on a composite network. The composite network in the method is composed of two or more sub networks and connecting edges between the sub networks. The sub networks are independent networks constituting the composite network. Thecomposite network includes an intersection network and a sensor network. The sensor network takes traffic sensors at intersections as nodes, and the sensors are related by connecting edges. The intersection network takes intersections as nodes, and connecting edges are established according to whether the intersections can be connected. On the basis of the intersection network, the definitions ofa congestion index, a congestion point, a congestion line and a congestion area are given. Calculation is carried out based on the network structure. The method specifically comprises the steps as follows: first, defining a congestion index Ti, a congestion point, a congestion line and a congestion area; then, quantifying the congestion level; and finally, analyzing and calculating congestion points, congestion lines and congestion areas. The method is designed ingeniously, has good operability and practicability, and has a wide market prospect.

The invention provides an identity authentication method and an identity authentication system based on a human face, which are used for solving the problem of distinguishing real persons from photos in the authentication process of the human face. The identity authentication method comprises the following steps of: carrying out human face authentication and facial expression recognition on a current frame image; carrying out comprehensive authentication on a human face authentication result and a facial expression recognition result of a continuous multiframe according to a preset strategy; and if the comprehensive authentication result is human face authentication pass and at least comprises two expressions, passing the identity authentication of the human face. According to the identity authentication method and the identity authentication system provided by the invention, by combining the expression recognition with human face recognition, the authentication performance is improved and the phenomenon that a jactitator passes the authentication by using a photo is avoided.

The invention belongs to the field of traffic condition prediction, and relates to a method for discovering a congestion time period based on a composite network. According to the method, a compositenetwork of an intersection network and a sensor network is constructed based on a complex network theory, and a congestion time period in the traffic network can be accurately judged. The method is implemented specifically according to the following steps: firstly, a congestion index Ti, a congestion time period and congestion level quantization are defined, then the congestion time period is calculated through establishing a mapping relation between the intersection network and the sensor network and performing congestion time period analysis, and finally frequent congestion time period analysis is performed. The model established according to the method is simple, the calculation data amount is small, the processing process is concise and clear, the operability and the practicability aregood, and effective implementation is facilitated. Meanwhile, the overall design is ingenious in conception, the required infrastructure are easy to obtain, the prediction process is simple, the application environment is friendly, and the market prospect is broad.

The invention discloses a denoising, feature extraction and pattern recognition method for human body surface electromyography signals, which comprises the following steps: performing best wavelet packet node denoising on the human body surface electromyography signals; performing wavelet decomposition on the human body surface electromyography signals after denoising, and extracting maximum absolute values of all levels of wavelet high-frequency coefficients as feature vectors; and inputting the extracted feature vectors into a back propagation network for performing training and pattern recognition. With the adoption of the best wavelet packet leaf node denoising method provided by the invention, noise carried in the electromyography signals can be effectively removed, and useful information can be retained; the maximum absolute values of the wavelet high-frequency coefficients can well reflect features of the human body surface electromyography signals, thereby having obvious advantages in comparison with the existing feature value method; and the relatively high recognition effect can be obtained by searching optimal hidden neurons and training errors of the back propagation network.

The invention discloses an automatic segmentation system for a digestive tract internal image. The system comprises an image preprocessing module, a feature extraction module and an image segmentation module, wherein the signal output end of the image preprocessing module is connected with the signal input end of the feature extraction module, and the signal output end of the feature extraction module is connected with the signal input end of the image segmentation module. Through the automatic segmentation system, the time for a doctor to search for images of different parts of the digestive tract can be shortened, the workload is reduced, and work efficiency is improved.

The invention discloses a visual simulation method for dynamic performance of a railway track system. The profile data of the steel rails and train wheels within the influencing range of train vibration are added with the vibration displacement data to rebuild steel rails and train wheels with a vibration shape, simulate deformation under force of steel rails and train wheels, thus realizing visual simulation of vibration behaviors of steel rails and train wheels in a railway track system. The method can more actually comprehensively simulate various dynamic performance of a track system while a train is running, and is characterized by high integrity; the invention can also visually display the dynamic status of the train and the track in a three-dimension scene, which is intuitionistic, vivid and simple for use.

The invention discloses a multipoint flatness assessment method based on support vector regression, wherein the method comprises the following steps of: sampling points on the surface of a measured flat plate to acquire a three-coordinate measured value of each point; pre-processing the original measured points, measuring a three-dimensional convex housing of a point set by calculating to remove measured points in the convex housing and keep the measuring points on the convex housing; using a support vector regression method-epsilon-SVR (Support Vector Regression) to acquire a least contain area plane of the measuring point set, wherein the corresponding contain area width is the acquired flatness. With the method, the calculation data quantity in the multipoint flatness assessment can bereduced, the calculating efficiency and the degree of accuracy of the assessment result can be improved.

The invention relates to the technical field of photographing measurement and disclsoes a photographing measurement method and device. The photographing measurement method comprises: acquiring image information; carrying out distortion correction on adjacent images; carrying out feature point extraction on the adjacent images subjected to the distortion correction; determining control points of the adjacent images; splicing the adjacent images according to the control points; adopting equal-distance cylinder projection to generate a spliced image; calibrating according to a camera and acquiring elements of interior orientation; according to a space resection equation, calculating elements of exterior orientation of each image; according to the elements of interior orientation, the elements of exterior orientation of the adjacent images and a space intersection equation, calculating ground coordinates; measuring the ground coordinates. The photographing measurement method and device have the advantages of small calculation data quantity, simple treatment and relatively high precision; the camera does not need to be rotated by a camera cloud deck; real three-dimensional coordinates of any measureable point in a panorama can be solved through one time of photographing and a point cloud is generated, so that measurement of distances, slopes, areas and the like can be carried out.

The invention discloses a multi-measuring-point planeness evaluation method based on support vector classification. The method comprises the following steps: sampling points on a tested surface to obtain a three-dimensional coordinate measurement value of each point; calculating a least square surface corresponding to a measurement point set; with the least square surface as a reference, removing measurement points with a small residual absolute value; moving measurement points with a big residual absolute value for the same distance respectively along the positive direction and the negative direction of the normal direction of the least square surface to form a linearly separable positive point set and negative point set; calculating the maximum interval plane of the positive point set and the negative point set with a linearly classified support vector machine method; and calculating a planeness value according to the plane. According to the multiple-measuring-point planeness evaluation method, the calculation data amount in the multiple-measuring-point planeness evaluation can be reduced and the calculation efficiency and the accuracy of an evaluation result are obviously improved.

The invention provides a voltage control method and device for a power distribution network provided with a distributed power source. The method includes the following steps that: the main nodes of each sub power distribution area in the power distribution network are obtained; the sensitivity matrix of the voltages of the main nodes and the power of the sub power distribution areas is determined;and a main node voltage allowable interval is determined, and voltage control is performed on the power distribution network provided with the distributed power source. According to the voltage control method and device of the invention, the optimized scheduling means of the distributed power source is provided; the flexible and efficient utilization of the distributed power source can be realized; a voltage quality requirement of the connection of a large-scale distributed power source into the power distribution network can be satisfied; the main nodes are adopted as the objects of voltagecontrol, so that the amount of calculation data on a global layer is decreased, and calculation efficiency can be effectively improved; and the frequent action of a control system can be avoided; andthe voltage control capability of the power distribution network with the distributed power source is improved; and the operational safety of the power distribution network with the distributed powersource is ensured.

The invention provides a space small file data distribution storage method and system based on access log information. The method includes: dividing a space small file data set into a frequently-accessed sub-set and a non-frequently-accessed sub-set, extracting the access sequence of the frequently-accessed space small file sub-set, calculating the association degree of each frequently-accessed space small file datum, and using the values of the association degrees to form an association matrix; performing magnitude conversion on each value in the association matrix, using an RCM sorting algorithm to rearrange the values, then outputting the values, using a local approximation search method to search for the optimal combination of the rearranged association matrix, using the optimal combination to perform distributed storage on the frequently-accessed space small data, and separately storing the non-frequently-accessed space small file data according to space position neighboring relations.

The invention relates to an electromagnetic spectrum map construction method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring monitoring data and a to-be-estimated point position, and acquiring path attenuation component estimation by using a trend surface fitting algorithm; subtracting the monitoring data from the path attenuation component estimation value of the corresponding position, performing position clustering on the obtained residual monitoring data, and estimating a clustering result in an optimal neighborhood by adopting a Kriging method to obtain shadow fading component estimation; and obtaining an estimation result according to the path attenuation component and the shadow fading component. Electromagnetic spectrum map construction is decomposed into estimation of path attenuation and shadow fading components, so that the electromagnetic spectrum map construction precision is effectively improved; through monitoring data clustering and optimal neighborhood selection, the calculated amount is reduced under the condition of ensuring the construction precision, and the electromagnetic spectrum map construction speed is effectively improved; and the method does not need prior information such as a radiation source and a propagation model, and has good adaptability.

The invention discloses a target detection method with visible light and far infrared rays fused. The method comprises steps: an acquired visible image and an acquired far infrared thermal imaging image are blocked, and the obtained image blocks are converted to N*1 vectors; a mixed sampling matrix is constructed to carry out acquisition and compression on converted vectors to obtain each image block after compression; a pixel-level background modeling algorithm is adopted to detect each pixel point in each image block after compression, and a target area and a background area in the image block are determined; and the detected target image and the far infrared thermal imaging image are overlapped to obtain a target image after fusion. As the mixed sampling matrix is adopted to sample and compress the image, the calculation data amount is reduced, the learning rate is adjusted according to an area, the average calculation time is reduced, the pixel number is reduced on the whole, the background modeling time is effectively reduced, and compared with the traditional target detection algorithm, the target detection method of the invention enables the memory capacity to be reduced by three fourths and the processing time can be reduced by more than 40%.

The invention relates to a picture similarity detecting method and system. The detecting method comprises a step of screening out Internet picture feature points in an Internet picture and reference picture feature points in a reference picture, and a step of obtaining matched feature points in the Internet picture and the reference picture according to all Internet picture feature points and allreference picture feature points, wherein the Internet picture and the reference picture are similar to each other when the number of matched feature points is larger than or equal to a preset threshold. According to the embodiment of the present invention, by obtaining and comparing the feature points of two pictures, whether the feature points of the Internet picture and the reference picture are matched is judged, the number of matched feature points is counted, when the number of matched feature points is larger than or equal to the preset threshold, a condition that the Internet picture and the reference picture are similar to each other is confirmed, in the embodiment of the present invention, the amount of calculated data is small, the anti-noise capability is strong, and whether pictures are similar to each other is quickly identified and checked.

A turning tool wear state monitoring method based on metric learning comprises the steps: preprocessing a collected tool wear image and then randomly dividing the image into a training set and a testset in proportion, wherein the training set and the test set are each divided into a support set and a test sample set; establishing a non-parameterized neural network model structure, wherein the non-parameterized neural network model structure comprises a feature extraction network, a distance metric function and an action mechanism; initializing weight, deviation and learning rate parameters ofthe feature extraction network model, and constructing an activation function; taking a cosine distance function as a distance metric function, wherein the Attention mechanism uses a softmax layer structure, and the Attention mechanism uses a softmax layer structure; calculating the current cross entropy loss by using forward propagation, and calculating the gradient of the current cross entropyloss through backward propagation; optimizing the weight and the deviation by operating a gradient descent method; and finally predicting, classifying and distinguishing the test set by using the learning parameters of the network model, and outputting the prediction accuracy. The method has the characteristics of simple algorithm implementation, error correction, high tool wear state classification accuracy, short operation time, strong real-time performance and the like.

The invention discloses a real-time vision system oriented target compression sensing method, which comprises the following steps: image reconstruction, mixed compression sensing, high-efficient Vibe target detection, updating and post-processing, wherein the step of the image reconstruction comprises the following specific steps: according to the size of a collected image, carrying out 4*4 partitioning on an image, and converting obtained image blocks into 16*1 vectors; the step of the mixed compression sensing comprises the following specific steps: constructing the image block corresponding to a mixed sampling matrix, and carrying out sampling compression; the step of the high-efficient Vibe target detection comprises the following specific steps: for each pixel in the image block, comparing a pixel value with a sample set to judge whether the pixel belongs to a background point or not; the step of updating comprises the following specific steps: according to the above detection result, determining a background block area and a target block area in the image block, and obtaining the parameter regulation information of the mixed sampling matrix of a next frame of image of the pixel according to a situation that the pixel belongs to the background block area or the target block area; and the step of the post-processing comprises the following specific steps: carrying out image optimization processing on each image block of the current frame to obtain a final target image of the current frame.

The invention discloses a sensing behavior recognition method and a sensing behavior recognition device. The sensing behavior recognition method comprises the following steps: acquiring sensing information acquired by an infrared sensing device, wherein the infrared sensing device consists of sensor arrays arranged at different acquisition points; according to the sensing information, establishing a target information matrix, wherein the target information matrix is used for showing sensing data determined by the sensor arrays at different time points; according to a deep convolution neural network algorithm, analyzing the target information matrix to determine a sensing behavior. Through the sensing behavior recognition method and the sensing behavior recognition device, small-calculation data amount, non-invasive, low-cost and small-volume body behavior recognition is achieved and the recognition accuracy rate is improved.