53 results about "Regression result" patented technology

A detection device has a neural network process section performing a neural network process using parameters to calculate and output a classification result and a regression result of each of frames in an input image. The classification result shows a presence of a person in the input image. The regression result shows a position of the person in the input image. The parameters are determined based on a learning process using a plurality of positive samples and negative samples. The positive samples have segments of a sample image containing at least a part of the person and a true value of the position of the person in the sample image. The negative samples have segments of the sample image containing no person.

The invention relates to a method for reconstructing an interval transit time curve by virtue of multiple logging curves. The method mainly comprises the steps of selecting a plurality of logging curves obvious in response to reservoir characteristics by performing reservoir sensitivity and correlation analysis on the logging curves; performing discrete wavelet decomposition on the logging curves together with a sonic wave curve to form 8 decomposition layers; enabling all the layers of high-frequency decomposition results of other logging curves except the sonic curve to form matrices respectively, obtaining the corresponding characteristic values and characteristic vectors of the matrices; taking the characteristic vectors corresponding to different characteristic values as new components, which, at the moment, have orthogonality (correlation ); performing multivariate regression analysis on the high-frequency components of the corresponding layer of wavelet decomposition of the sonic curve by utilizing characteristic vectors of each layer, calculating a weighting coefficient corresponding to each vector, returning a regression significance analysis result and determining the quality of regression; next, determining the number of layers of the sonic high-frequency components to be reconstructed by virtue of the characteristic vectors according to the regression significance analysis result, selecting multiple regression results as the high-frequency components for a part having more layers than selected layers, and remaining the high-frequency decomposition results of the sonic logging curves for a part having less layers than the selected layer, and then carrying out curve reconstruction by using the low-frequency components of the sonic logging curves and the high-frequency components obtained through regression so as to obtain a final sonic reconstructed curve.

The invention discloses an electric transmission line insulator burst identification method based on deep learning. The method comprises the following steps that: firstly, collecting an electric transmission line insulator picture by an unmanned aerial vehicle; then, carrying out target detection on the electric transmission line insulator picture collected by the unmanned aerial vehicle so as toaccurately carry out regression to obtain the position of an insulator string in an original picture, and clipping an independent insulator string according to a position regression result; then, adopting a deep learning full-conventional neural network method to carry out semantic segmentation on the insulator string, and segmenting the insulator string from a background; and finally, extractingthe center of mass of the single insulator, solving a distance between the centers of mass of adjacent insulators, setting 1.5 times of average distance as a threshold value, considering that burst isin the presence in two insulators if the distance between two insulators is greater than the threshold value, and labeling an insulator burst position. By use of the method, a subjective influence brought by manually setting the threshold value and selecting the parameter in a traditional insulation extraction process is avoided, and identification accuracy is improved.

The invention provides an aesthetic attribute evaluation method based on a dense convolution network and a multi-task network. Aesthetic features of images are extracted from dense convolution neuralnetwork model, and a part of image features are preserved in multi-dimensional matrix. The hierarchical multi-task network is used for regression analysis of known image attributes. After many times of training analysis, the prediction results and the data in the training data set reach a high degree of fitting, the final training model is saved, and the model is tested on the test data set to getthe regression results of the method. Because the data set used does not have a certain tendency, the aesthetic attribute prediction algorithm model has a certain universality. The method is implemented using Google's Tensorflow framework, and can be widely used in computer vision, image analysis and processing, digital photography and digital entertainment and other fields.

The invention relates to a fast pedestrian detection method for complex public scenes based on depth learning, wherein the method includes pixel size preprocessing of a training image and a test image, pre-training of a convolution neural network based on classification tasks, pedestrian detection training of the convolution neural network based on pedestrian detection tasks, use of threshold filter to eliminate prediction boxes with low confidence, and use of non-maximum inhibition to eliminate multiple prediction of the same pedestrian. Cross entropy is used as loss function in pre-training.Finally, the improved mean square error is used as the loss function to make the network output the regression results of predicting the location of pedestrians. In the testing phase, the image is used as the input of the convolution neural network, and the threshold filter and non-maximum suppression are used to filter all the output prediction results of the convolution neural network, so thatthe pedestrian location information can be detected and the pedestrian intelligent monitoring can be realized.

The invention provides a performance continuous integration data processing method and device. The method includes acquiring a performance continuous integration test result document stored in a disk, wherein the performance continuous integration test result document comprises a plurality of performance continuous integration test records; reading the plurality of performance continuous integration test records into a memory one by one; analyzing the performance continuous integration test records in the memory to obtain corresponding middle performance index data; deleting the analyzed performance continuous integration test records in the memory; judging whether analyzing of the plurality of performance continuous integration test records in the performance continuous integration test result document is completed; if analyzing of the plurality of performance continuous integration test records in the performance continuous integration test result document is completed, acquiring final performance index data and deleting the performance continuous integration test result document in the disk; storing the final performance index data to the memory and the disk. According to the performance continuous integration data processing method and device, space usage of the disk can be reduced, and analyzing and reading efficiency of regression result data is improved.

The invention provides a multi-region people counting method based on a single camera, and the method comprises the steps: S1, collecting an image through the camera, and carrying out the preprocessing of the image; S2, respectively sending the preprocessed images to a detection network and a regression network; and S3, carrying out region division on the camera picture, and carrying out region judgment on the position of each target. When the method is applied to a real application scene, a detection result and a regression result are verified and judged, the real-time number of people in multiple areas in a single picture can be accurately and reasonably output in real time. The number of people in the whole picture is monitored, and meanwhile, a plurality of different specific areas aresupported to be specified, so that the time cost and the labor cost of statistics and monitoring are saved, the trouble caused by frequent statistics of the number of people in real time is avoided,and the unnecessary loss and the possibility of accidents are reduced.

The invention relates to the field of medical detection and particularly discloses a noninvasive prenatal biological information detection and analysis method. For improving the accuracy of analyzing different quantities of to-be-detected samples, different detection and analysis methods are selected according to the different quantities of the to-be-detected samples, and different analysis policies are adopted for parameters obtained by the to-be-detected samples and parameters obtained by a normal reference set, so that the accuracy of analysis is improved to a greater extent. According to the method, the problem of inaccurate regression result caused by great influence of abnormal data on slope due to use of a least square method for regression in a process of correction by using a whole chromosome method in the prior art is well solved by adopting robust regression and CV regression, so that the robustness and accuracy of sample analysis are ensured. A set of analysis method for judging anomaly of sex chromosome by utilizing a ZZ value is originated; and the chromosome anomaly is judged by using the ZZ value method, so that related statistic judgment standards are better met, a result is more accurate, and the reliability of the method for judging the anomaly of the sex chromosome is enhanced.

The invention provides an automatic remote sensing image haze detection method. The method comprises the steps: taking images without clouds as the reference, wherein the images are acquired in different time; establishing an HOT preliminary cloud detection result, a multi-element regression result between the HOT and a difference image between an image without clouds and an image with clouds, and a one more multi-element regression result based on the above results and the image with clouds; performing repeated iteration of the above processes; and then obtaining the final result for haze detection. The automatic remote sensing image haze detection method effectively overcomes the problem that traditional HOT (Haze Optimized Transformation) and BSHTI (Background Suppressed Haze Thickness Index) cloud thickness detection methods cannot suppress the noise of highlighted ground features, and especially can effectively distinguish ice and snow from haze, and provides an effective method for detecting the haze thickness for large amount of cloud contamination remote-sensing images.

The invention discloses a data classification regression method and a data classification regression device. The method comprises the following steps of dividing an initial sample vector set into a continuous type data sequence, a category type data sequence and a binary data sequence; respectively converting the continuous type data sequence and the category type data sequence into a first vector sequence and a second vector sequence which are in a binary form; merging the first vector sequence, the second vector sequence and the binary data sequence to generate a classification regression vector sequence; and obtaining a classification regression result of the initial sample vector set according to each vector in the classification regression vector sequence. By adopting the technical scheme, the obtaining of the data classification regression result is not limited by the data dimension number and the data volume; and the data classification regression can be realized without adopting an iterative algorithm, so that the complexity for obtaining the data classification regression result is lowered.

The invention discloses a rotation invariant face detection method based on a multi-task progressive registration network, and belongs to the field of computer vision. The method mainly comprises thefollowing steps: preprocessing an image, and constructing and training a cascaded multilayer convolutional neural network; inputting a test image, generating image sets with different resolutions by using an image pyramid mode, and then sending the image sets into the cascaded multilayer convolutional neural network to start detection; filtering out a part of non-face windows by each level of network, adjusting the position of a candidate frame according to a frame regression result, and predicting the rotation angle of the face at the same time; and then carrying out registration through image overturning operation according to the predicted rotation angle. According to the invention, through a multi-task progressive registration network method, real-time and rotary self-adaptive face detection is realized, and good effects are achieved in precision and speed.

The invention relates to a subgrade settlement prediction method based on an improved Verhulst curve. The method comprises the following steps: firstly, adding one index to a time factor in a Verhulst curve model to improve a Verhulst curve model, adjusting a time factor index when the improved Verhulst curve model is dissolved, and performing settlement prediction through the improved Verhulst curve model. By adopting the method, a subgrade settlement prediction value is relatively and highly related to real settlement data, a relatively satisfactory final settlement regression result is acquired, and a prediction method which is wide in applicability and high in reliability is provided for subgrade settlement monitoring in geotechnical engineering.

The invention discloses a multifunctional high-efficiency dynamic chip verification simulation method and device, which maintain a unified and configurable verification case management list, dynamically adjust the simulation execution quantity of verification cases through real-time interaction with a server cluster, and dynamically complete simulation operation of large-scale verification cases to the maximum extent in parallel. And after the regression is completed, secondary regression of the debugging mode is carried out on a regression result error use case. According to the method, hardware resources of the server cluster are fully utilized, the number of verification cases which are executed in parallel is dynamically adjusted in real time, the automation level of regression of theverification cases is improved, the workload of verification personnel is reduced, the research and development period of chip verification work is shortened, and the efficiency is more remarkably improved for the application condition of a multi-project shared server cluster.

The invention provides a cement raw meal three moduli measuring method based on partial least squares, relating to a method for detecting material components by laser induced plasma spectroscopy. The method comprises the following steps: firstly detecting cement raw meal calibration samples with known components by adopting an LIBS system and building calibration models of element concentration ratio and characteristic spectrum line intensity ratio by utilizing a spectrum standardization method according to the obtained characteristic spectrum line intensity; then carrying out online detection on the samples to be detected and obtaining LIBS spectra, judging the element concentration ratio through the regression results and then determining the three moduli of cement raw meal by substituting the spectrum line information obtained through measurement into the built calibration models. The method has the effects of taking full advantage of the spectrum line information obtained by the LIBS spectra, considering the influence of the interference elements on the measuring results and reducing the interference of matrix effects and has the characteristics of good goodness of fit, strong repeatability and high prediction accuracy compared with traditional univariate calibration methods.

The invention relates to an algorithm model operation monitoring method and device, computer equipment and a storage medium. The method comprises the steps of obtaining a feature tag and a regression result of an algorithm model in a current statistical period; determining a current performance index according to the feature tag and the regression result, wherein the current performance index is a performance index value of the algorithm model in a current statistical period; and when the current performance index value satisfies a preset alarm condition of the algorithm model, outputting first alarm prompt information. By adopting the method, the abnormal operation condition of the algorithm model can be found in time, the blindness of updating the algorithm model can be avoided, and the performance of the algorithm model can be ensured.

The invention provides a quick three-dimensional mask diffraction near-field calculation method based on sample library and data fitting. The method includes following particular steps: a) establishing a three-dimensional mask diffraction matrix sample library, and calculating diffraction near-field data correction factors corresponding to convex angles, concave angles and edge zones; b) to one three-dimensional mask requiring diffraction near-field calculation, determining an plurality of observation points on the mask, and distributing a sub zone to each observation point; (c) with each observation point as a center, setting a squared zone surrounding the observation point on the mask; (d) according to the squared zone and the sample library, with kernel regression technology and a data fitting method, respectively calculating a diffraction matrix regression result corresponding to each observation point; and (e) finally filling the regression results of all observation points into corresponding sub zones, thereby splicing the diffraction matrix regression result corresponding to the whole three-dimensional mask. The method considers the influence of corner structures to the diffraction near field in the three-dimensional mask graph, so that calculation precision of three-dimensional mask diffraction near-field calculation is increased.

The invention provides a target tracking method and device. The target tracking method comprises the following steps: acquiring a to-be-tracked image; inputting the to-be-tracked image into a trained target tracking model to obtain a target tracking result of the to-be-tracked image; wherein the target tracking model comprises a classification branch and a regression branch, the target tracking model is obtained by training a preset neural network based on a regression sample and a classification sample, and the regression sample is determined according to initial classification loss of an initial sample obtained through the classification branch; the classification sample is determined according to the initial regression loss of the initial sample obtained through the regression branch. The method effectively guarantees the alignment of classification and regression results in the training process, thereby improving the robustness and precision of the target tracking model, and effectively improving the efficiency and accuracy of target tracking by using the target tracking model.

The invention provides a medium-speed coal mill fault early warning system based on a least square support vector machine algorithm, and the system comprises the steps: (1) collecting historical data of a coal mill, preprocessing of the data, and selecting a parameter which is greatly related to a target regression parameter according to a Pearson's correlation coefficient; (2) inputting selected historical parameters into an LS-SVM algorithm for training to obtain an optimal regression result model; and (3) inputting actual parameters into the trained algorithm to obtain an actual value regression result, calculating an adaptive threshold interval, judging whether the actual value is in the adaptive threshold interval or not, considering that the parameters are abnormal if the actual value exceeds the adaptive threshold interval for continuous 15 seconds, and performing early warning on an abnormal state. The method can predict different faults according to different parameters.

The invention provides a wind power curve fitting method based on sparse heteroscedasticity multi-strip regression, and the method comprises the steps: automatically detecting an abnormal point through employing a fuzzy C-means algorithm, and obtaining the data of which the abnormal point is removed for original wind power data; constructing a sparse heteroscedasticity multi-strip regression modelaccording to the obtained data; optimizing the constructed sparse heteroscedasticity multi-strip regression model by adopting a variational Bayesian method to obtain posteriori distribution conditions and parameter formulas of all parameters in the model; and initializing model parameters, and solving estimated values of the parameters by utilizing an iterative method according to posteriori distribution conditions and parameter formulas of all the parameters in the model. According to the wind power curve fitting method based on sparse heteroscedasticity multi-spline regression provided by the invention, a plurality of spline basis functions is integrated, the nonlinear fitting capability of the model is improved, and the influence of redundant information on a final regression result isavoided.

The invention discloses a pedestrian re-identification method based on a sparse attention network. The pedestrian re-identification method comprises the following steps: firstly, transmitting shallowfeatures to deep features in a lossless manner through short connection; secondly, extracting a main volume and features of the image through a trunk residual network formed by continuously superposedresidual modules; extracting detail features, which are liable to be lost, of the image through a normalized compression-excitation module embedded in the backbone residual network; and finally, multiplying the obtained features, adding the features obtained in the first part, and conveying the features into a full connection layer and a classification regression layer to obtain classification and regression results. The sparse attention network can effectively extract pedestrian photo detail features of a plurality of pedestrian re-identification data sets.